Awuah et al. European Journal of Medical Research (2025) 30:61 https://doi.org/10.1186/s40001-025-02339-3 REVIEW Predicting survival in malignant glioma using artificial intelligence Wireko Andrew Awuah1*†, Adam Ben‑Jaafar2†, Subham Roy3, Princess Afia Nkrumah‑Boateng4, Joecelyn Kirani Tan5, Toufik Abdul‑Rahman1 and Oday Atallah6  Abstract  Malignant gliomas, including glioblastoma, are amongst the most aggressive primary brain tumours, character‑ ised by rapid progression and a poor prognosis. Survival analysis is an essential aspect of glioma management and research, as most studies use time-to-event outcomes to assess overall survival (OS) and progression-free survival (PFS) as key measures to evaluate patients. However, predicting survival using traditional methods such as the Kaplan–Meier estimator and the Cox Proportional Hazards (CPH) model has faced many challenges and inac‑ curacies. Recently, advances in artificial intelligence (AI), including machine learning (ML) and deep learning (DL), have enabled significant improvements in survival prediction for glioma patients by integrating multimodal data such as imaging, clinical parameters and molecular biomarkers. This study highlights the comparative effectiveness of imaging-based, non-imaging and combined AI models. Imaging models excel at identifying tumour-specific features through radiomics, achieving high predictive accuracy. Non-imaging approaches also excel in utilising clinical and genetic data to provide complementary insights, whilst combined methods integrate multiple data modalities and have the greatest potential for accurate survival prediction. Limitations include data heterogeneity, interpret‑ ability challenges and computational demands, particularly in resource-limited settings. Solutions such as federated learning, lightweight AI models and explainable AI frameworks are proposed to overcome these barriers. Ultimately, the integration of advanced AI techniques promises to transform glioma management by enabling personalised treatment strategies and improved prognostic accuracy. Keywords  Malignant glioma, Artificial intelligence (AI), Machine learning (ML), Deep learning (DL), Survival prediction approaches Introduction Malignant gliomas, including glioblastoma, are amongst the most aggressive primary brain tumours, known for their rapid progression and poor prognosis. Glioblas- toma, the most common subtype, has a dismal 5-year survival rate of only about 5.6% [1]. Several studies suggest that overall survival in glioblastoma patients is closely linked to the tumour’s anatomical location within the brain [2, 3]. A recent study by Osadebey et al. (2023) found that gliomas located in the hippocampus, thalamus, left insula and regions of the left lateral ven- tricle were associated with short survival [2] Gliomas in the frontal and temporal lobes are associated with Open Access © The Author(s) 2025. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by-​nc-​nd/4.​0/. European Journal of Medical Research †Wireko Andrew Awuah and Adam Ben-Jaafar contributed equally and are co-first authors. *Correspondence: Wireko Andrew Awuah andyvans36@yahoo.com 1 Department of Research, Toufik’s World Medical Association, Sumy, Ukraine 2 School of Medicine, University College Dublin, Belfield, Dublin 4, Ireland 3 Hull York Medical School, University of York, York, UK 4 University of Ghana Medical School, Accra, Ghana 5 Faculty of Biology, Medicine and Health, University of Manchester, Manchester M13 9PL, UK 6 Department of Neurosurgery, Carl Von Ossietzky University Oldenburg, Oldenburg, Germany http://creativecommons.org/licenses/by-nc-nd/4.0/ http://crossmark.crossref.org/dialog/?doi=10.1186/s40001-025-02339-3&domain=pdf Page 2 of 11Awuah et al. European Journal of Medical Research (2025) 30:61 intermediate survival, whilst tumours in the corpus callo- sum, right insula and regions of the right lateral ventricle are associated with long survival [2]. Other studies have also reported that central tumour location is associated with short survival, whilst survival is favourable accord- ing to the distance between the centre of the third ven- tricle and the contrast-enhancing tumour margin [3]. In addition, patients with gliomas in non-eloquent areas of the brain have been found to have favourable survival compared to patients with tumours in eloquent or near- eloquent areas, irrespective of the extent of resection [4]. The management of gliomas has been challenging for decades due to difficulties in accurate diagnosis and tai- lored treatment. Early detection and accurate assessment of tumour progression are hampered by tumour hetero- geneity, the variability of molecular markers and the limi- tations of current imaging techniques [5]. In addition, the highly infiltrative nature of gliomas, combined with the delicate structure of the brain, makes complete sur- gical resection difficult [6]. Standard treatment typically involves maximally safe surgical resection followed by radiotherapy and concurrent chemotherapy with temo- zolomide [7, 8]. However, despite these extensive inter- ventions, median survival remains around 14  months, with progression-free survival (PFS) often limited to a few months [8]. Survival analysis is essential in clinical neuro-oncology research, as most studies use time-to-event outcomes to evaluate overall survival (OS) and progression-free sur- vival (PFS) as key measures to assess patient prognosis after cancer diagnosis or recurrence [9]. However, the complex nature of gliomas has made accurate predic- tion of OS a major challenge for clinicians for decades. Recent developments have led to new technologies that can improve the accuracy of OS prediction, help- ing physicians to develop more comprehensive, person- alised treatment plans that are best suited to individual patients. Historically, glioma survival has been predicted using traditional methods such as the Kaplan–Meier esti- mator and the Cox Proportional Hazards (CPH) model, which estimate survival probabilities and take into account prognostic factors [10]. Despite their widespread use, these methods have notable limitations, including the proportional hazards assumption and reduced power when applied to high-dimensional data, such as molecu- lar biomarkers or complex imaging features [9]. In addi- tion, these models struggle with non-linear interactions between variables and often require prior knowledge of the factors influencing outcomes [11]. Recent advances in artificial intelligence (AI) show promising potential to address these challenges by integrating complex data- sets and generating personalised survival predictions. AI can handle large datasets, capture non-linear patterns and provide individualised risk assessments, achieving greater accuracy than traditional models [12]. This review aims to evaluate the application of AI and machine learning (ML) in predicting survival outcomes for glioma patients, to assess their performance relative to traditional statistical methods, and to explore their potential to improve clinical decision making in glioma management. Methods This narrative review aims to establish a comprehensive framework for predicting survival in patients with malig- nant glioma using AI. To enhance methodological rigour, a comprehensive selection process was employed based on specific inclusion and exclusion criteria. Only full-text articles published in English were included, and searches were conducted in several major databases, including PubMed/Medline, EMBASE, the Cochrane Library and Scopus. A wide range of targeted keywords such as "malignant glioma", "glioma", "glioblas- toma", "brain tumour", "malignant brain tumour", "arti- ficial intelligence", "machine learning", "deep learning", "imaging AI models", "non imaging AI models", "deep learning models", "machine learning models", "convolu- tional neural networks", "survival prediction approaches", "predictive modelling", "survival analysis", "non-multi- modal neuroimaging", "multimodal imaging", "combined imaging", "MRI scan", "CT scan", "multimodal MRI", "multimodal PET" and "non-multimodal MRI" guided an exhaustive database search. In addition, references from recent reviews on related topics were manually screened to identify additional sources that could enrich the search strategy. The review included studies published between 2004 and 2024 to cover two decades of research progress on the topic. Studies included descriptive, preclinical/ animal model, cohort and observational research from clinical settings to provide a multidimensional under- standing of the topic. Exclusion criteria included stand-alone abstracts, con- ference proceedings, letters to the editor, editorials, per- spectives and posters to focus on high quality and reliable studies. Studies that were not peer reviewed and not pub- lished in English were also excluded. A summary of the methodology, including inclusion and exclusion criteria, is provided in Table 1. Categories of survival prediction models for malignant gliomas Malignant glioma survival prediction models use differ- ent types of data, including imaging and non-imaging data such as genomic and clinical parameters, to improve prediction accuracy. The following subsections catego- rise the types of models into ML, a subset of AI, Deep Page 3 of 11Awuah et al. European Journal of Medical Research (2025) 30:61 Learning (DL), a subset of ML, and the statistical model, whilst briefly discussing how these methods thrive using imaging, non-imaging, and combined methods. The suc- cess of ML, DL and statistical models in predicting gli- oma OS is largely dependent on the integration of both imaging and non-imaging data. ML models excel at com- bining features from both MRI scans and clinical data, DL models excel with high-dimensional imaging data, and statistical models also excel at providing interpret- ability of survival probabilities [10, 13, 14]. Using these methods in combination, predictive models for glioma survival become more accurate and personalised, provid- ing valuable insights for clinical decision making. ML‑based survival prediction models ML-based survival prediction models for malignant glio- mas use both imaging and non-imaging data to improve prediction accuracy. Imaging-based models often use magnetic resonance imaging (MRI) scans, which can be non-multimodal (using a single type of imaging) or mul- timodal (combining multiple types of imaging, such as structural MRI and functional MRI). For example, recent systematic analyses have shown that the most com- monly used ML algorithms are support vector machines (SVMs), random survival forests (RSFs), boosted tree methods and artificial neural networks (ANNs) [9, 13]. These algorithms have been successfully applied to radio- logical images, particularly MRI, to predict survival out- comes by extracting features such as tumour volume and shape [15]. Non-imaging models typically rely on clinical data such as patient age, genetic markers and treatment information to estimate survival. ML methods excel at integrating these features, improving the robustness of predictions and providing actionable insights in clinical settings [12]. Furthermore, models that combine imaging and non-imaging data have shown the best performance, as they integrate structural, functional and clinical fea- tures to provide personalised survival predictions [1]. Recent studies highlight how ML thrives when combin- ing radiomics (from MRI scans) with molecular biomark- ers or clinical data. The integration of radiomic features with non-imaging data in ML models has significantly improved the accuracy of predicting PFS in gliomas. By capturing complex interactions between diverse data types, ML methods demonstrate a high degree of adapt- ability to multimodal inputs, further enhancing their prognostic capabilities [6, 9]. ML models like RSF have also shown promise in predicting survival outcomes by handling mixed data types effectively [11]. However, ML methods are prone to overfitting, particularly when applied to small or unbalanced datasets, which compli- cates the extraction of clinically relevant information. Additionally, addressing challenges such as missing data and the integration of multimodal datasets—which includes clinical, imaging and molecular data—often necessitates the use of advanced preprocessing tech- niques [13]. DL models for glioma survival prediction DL models for glioma survival prediction rely heav- ily on imaging data. For example, convolutional neural networks (CNNs) can automatically learn features from structural MRI and multimodal imaging [16]. These DL methods are effective at integrating different types of data—imaging, genomic and clinical—and enable more accurate and reliable survival predictions by captur- ing complex patterns and relationships across modali- ties [13]. These DL approaches have demonstrated Table 1  Summary of methodology Methodology steps Description Literature search PubMed/MEDLINE, EMBASE, Scopus and the Cochrane Library Inclusion criteria Various study designs including experimental studies, randomised controlled trials, prospective and retrospective cohort studies Studies involving both paediatric and adult populations Studies providing raw data Full-text articles published in English Exclusion criteria Non-English and non-peer reviewed studies, stand-alone abstracts, conference proceedings, editorials, commentaries, and letters Search terms "malignant glioma", "glioma", "glioblastoma", "brain tumour", "malignant brain tumour", "artificial intelligence", "machine learning", "deep learning", "imaging AI models", "non imaging AI models", "deep learning models", "machine learning models", "convolutional neural networks", "survival prediction approaches", "predictive modelling", "survival analysis", "non-multimodal neuroimaging", "multimodal imaging", "combined imaging", "MRI scan", "CT scan", "multimodal MRI", "multimodal PET" and "non-multimodal MRI" Additional search A manual search was performed to include references from recently published procedure-specific and disease-specific reviews Sample size requirement No strict sample size requirement Page 4 of 11Awuah et al. European Journal of Medical Research (2025) 30:61 remarkable performance, particularly with high-dimen- sional imaging data, by recognising intricate patterns associated with glioma progression. Whilst DL models thrive primarily on imaging data, they can also incor- porate non-imaging features, such as clinical or genetic data, to further improve survival predictions. DL-based models excel at integrating both imaging and clinical information, providing deeper insights into glioma char- acteristics and prognosis [17]. Holistic models that combine imaging with molecu- lar data have led to improved predictive accuracy for glioblastoma survival [18]. For example, the integration of positron emission tomography (PET) imaging with histopathological features allows DL models to cap- ture both physiological and molecular characteristics of tumours, providing more robust survival estimates [19]. In addition, advanced DL models, such as 3D CNNs, have proven particularly effective in identifying key brain regions that influence survival, providing clinicians with interpretable results [14]. Statistical models Statistical models, such as CPH and Kaplan–Meier esti- mates, have traditionally been used for survival analy- sis in gliomas. However, these models are often limited when it comes to handling complex, high-dimensional data. Despite this, they remain valuable for integrating clinical data (e.g., demographic features and treatment history) to generate survival estimates. Statistical meth- ods, when combined with imaging or molecular data, can complement ML and DL models by adding interpret- ability and robustness to survival predictions [10]. These models are also useful in scenarios where simpler, more interpretable models are preferred over black-box meth- ods like DL. Recent advancements have demonstrated that integrat- ing statistical approaches with ML and DL can enhance survival predictions. For example, hybrid models that combine statistical techniques with machine learning methods like support vector regression have provided better predictive performance by leveraging both clini- cal data and imaging features [20]. Moreover, statistical models like nomograms that incorporate both clinical and imaging data offer a more interpretable means of predicting survival, facilitating personalised treatment strategies [21]. Table  2 summarises the types of AI sur- vival prediction models used for malignant gliomas. Survival predictions and outcomes based on imaging, non‐imaging and combined AI models Imaging AI models AI-based survival prediction in patients with malig- nant gliomas has leveraged both imaging techniques; non-multimodal and multimodal approaches that inte- grate clinical and molecular data. Non-multimodal imaging methods, such as MRI and CT, have provided significant insights into patient prog- nosis. For example, a model using only MRI data from glioblastoma patients achieved a C-index of 0.78 by strat- ifying patients based on radiomic features from T1- and T2-weighted images [22]. MRI-derived radiomics has demonstrated an 83% accuracy in identifying biomark- ers, such as tumour shape and texture, that correlate with survival [18]. Furthermore, the utilisation of post-radio- therapy MRI data achieved a high area under the curve (AUC) of 0.93 in predicting survival in glioblastoma patients, underscoring the potential of capturing imag- ing data at different treatment stages [23]. Similarly, CT- based models, with a C-index of 0.74, have highlighted tumour volume and enhancement patterns as key pre- dictors of survival [17]. DL applied to histopathological images of glioma tissue achieved 87% accuracy, identify- ing nuclear pleomorphism and mitotic activity as indi- cators of poor prognosis [24]. Meanwhile, a DL model that focused on glioblastoma features like necrosis and oedema achieved 81% accuracy, strongly correlating with patient outcomes [2]. Furthermore, quantitative features from MRI scans can classify glioma patients into survival categories with up to 98% accuracy [25]. Non-multimodal scans provide detailed insights into tumour volume, texture, and intensity, beyond what clinical data alone can offer. These models provide a non-invasive alternative for prognostication, enabling cli- nicians to predict outcomes and tailor treatments effec- tively. Their high accuracy and specificity illustrate their potential to replace invasive methods, paving the way for broader applications in neuro-oncology. However, their reliance solely on imaging data limits their ability to cap- ture systemic and molecular-level nuances that signifi- cantly influence survival [26, 27]. Multimodal imaging, which combines MRI, CT, and PET, leverages the complementary strengths of radi- omic features and clinical or molecular information. Multimodal AI overcomes the limitations of single- modality approaches by providing greater accuracy and reliability, enabling comprehensive survival stratifica- tion and improved personalised treatment strategies. This integrated approach addresses the complexities of glioblastoma prognosis and represents a significant step forward in patient care [28, 29]. This extensive approach has further enhanced predictive accuracy. For example, an ML model applied to multiple imaging modalities achieved approximately 82% accuracy, outperforming traditional methods [20], whilst DL models integrating MRI and CT data have also achieved over 85% accuracy by identifying complex patterns across imaging types Page 5 of 11Awuah et al. European Journal of Medical Research (2025) 30:61 [30]. An online survival prediction tool using multi- modal imaging with WHO CNS5 data, along with mod- els that integrate molecular and clinical information, has achieved a C-index of 0.75 [12]. In addition, quantitative radiomics features reflect- ing tumour heterogeneity and phenotype are extracted from imaging modalities such as multi-parametric MRI (mpMRI). Radiomics signatures derived from mpMRI are shown to stratify glioblastoma patients into survival groups with high accuracy. Imaging features have been integrated with clinical variables to further improve sur- vival prediction using ML classifiers, such as ensemble learning models [31]. Imaging-based survival prediction has also ben- efited from DL techniques. Studies have proposed a multi-channel 3D DL architecture using multimodal neuroimaging data. The framework was built using contrast-enhanced T1 MRI, diffusion tensor imaging (DTI) and resting-state functional MRI (rs-fMRI) and achieved an accuracy of 90.66% in classifying survival outcomes. Such studies demonstrate the promise of DL for interpreting complex imaging datasets to improve clinical decision making [32]. Radiomics extracted from preoperative contrast- enhanced MRI combined with linear discriminant analysis (LDA) had high predictive accuracy for 3- and 6-month survival in glioblastoma patients, with AUC values of 0.88 and 0.78, respectively [26]. Similarly, automated glioma grading using CNNs with high sen- sitivity and specificity allows survival stratification without invasive biopsy. These image-based approaches exploit spatial and textural information, laying the foundation for AI in non-invasive survival prediction [26, 27]. Whilst accurate and specific, imaging models may miss critical non-imaging factors such as genetic mutations or clinical history. Therefore, a combined approach that integrates imaging and non-imaging data is more prom- ising, providing a more comprehensive framework for survival prediction and personalised treatment strategies. Table 2  Types of survival prediction models for malignant gliomas MRI, Magnetic Resonance Imaging, fMRI, Functional Magnetic Resonance Imaging, PET, Positron Emission Tomography, ML, Machine Learning, 3D, 3-Dimensional, DL, Deep Learning, CNN, Convolutional Neural Networks, CPH, Cox Proportional Hazards, PFS, Progression-free Survival Prediction model type Data types Methodology Key features/outcomes ML model  Dynamic nomograms [21] Molecular biomarkers, clinical parameters Statistical modelling Enables individualised predictions for personalised treatment decisions  Non-imaging models [12] Clinical and molecular data ML (CPH, Support Vector Machines) Provides robust survival estimates through integration of demographic and clinical features  Radiomics-based methods [6, 9] MRI-derived imaging features, clinical data ML techniques Achieves high accuracy in predicting PFS using feature extraction DL model  Imaging-based models [20] MRI scans (structural, functional) Neural networks Utilises multimodal approaches com‑ bining resting-state fMRI and struc‑ tural MRI  Predictive performance models [15] Radiology, pathology imaging Ensemble regression, DL Enhances predictive performance by integrating radiology and pathol‑ ogy images  Holistic models [19] In vivo PET imaging, ex vivo histo‑ pathology Integrated modelling Captures physiological and molecular characteristics for improved survival predictions  3D convolutional neural net‑ works [2] MRI scans, clinical data DL (3D CNNs) Provides interpretable outputs highlighting critical brain regions influencing survival predictions Statistical model  Traditional models (e.g. CPH, Kaplan–Meier estimates) [10] Clinical data (demographics, treat‑ ment history) Statistical survival analysis Provides interpretable survival estimates but struggles with high- dimensional data  Hybrid statistical models (e.g. CPH combined with ML) [11] Clinical and imaging data Combination of CPH and ML techniques Enhances predictive performance by integrating interpretability with non-linear data analysis  Nomograms [21] Clinical and imaging data Statistical modelling Facilitates personalised predictions with interpretable survival prob‑ abilities Page 6 of 11Awuah et al. European Journal of Medical Research (2025) 30:61 Non‑imaging AI models AI applications for predicting survival outcomes in malignant gliomas are increasingly utilising non-imaging approaches, focusing on clinical data and histopathologi- cal features. For instance, an online tool that combines traditional statistical methods with ML uses variables such as Karnofsky Performance Status (KPS) and patient demographics to improve survival predictions in glio- blastoma patients [33]. The inclusion of inflammatory biomarkers has also proven beneficial, with models incorporating these markers achieving a C-index of 0.78, surpassing the accuracy of traditional methods [34]. Furthermore, inte- grating quality-of-life assessments into survival mod- els has reduced the mean absolute error to 3.4  months, demonstrating the value of patient-reported outcomes in enhancing predictive accuracy [35]. Alternative stud- ies have achieved similar results using WHO CNS5 data, demonstrating high predictive accuracy with AUC val- ues of 0.849, 0.835, and 0.821 for 1, 3, and 5-year over- all survival predictions, and identified key prognostic factors like age, IDH1, and CDKN2A alterations [21]. Additionally, when clinical and demographic data were incorporated into a DL model, it achieved 85% accuracy in predicting 1-year survival outcomes [36]. Whilst imaging remains critical, non-imaging data such as genetic, molecular and clinical parameters pro- vide complementary insights. Studies have highlighted the importance of multi-type genetic data, including mRNA expression, DNA methylation and microRNA profiles, to address cancer heterogeneity. A DL approach effectively captured common and specific genetic fea- tures and outperformed conventional methods in sur- vival prediction accuracy [37]. Similar studies have also investigated the use of ML techniques such as ANNs and SVMs in analysing small, heterogeneous glioma datasets. The results showed that these techniques outperformed traditional statistical methods, and that the inclusion of demographic and clinical variables was critical for more nuanced survival predictions. Such non-imaging approaches highlight the importance of AI in harnessing diverse data to accurately predict patient outcomes [38]. Transformer-based models have been introduced for glioblastoma survival prediction by integrating clinical and molecular pathology data. These models achieved consistent performance across multiple datasets, high- lighting their generalisability and reliability. Using high- dimensional data integration, this approach provides insights into survival determinants beyond anatomical imaging [14]. Another important development is the use of ML algorithms to investigate non-imaging variables such as age, performance status and genetic mutations. The combination of clinical features such as Ki-67 and P53 mutation status with ML algorithms improves sur- vival prediction beyond traditional statistical methods [39]. AI-driven non-imaging approaches using clinical, genetic and molecular data improve glioma survival pre- diction. Techniques such as ML, ANN and transformers outperform traditional methods by integrating variables such as KPS, biomarkers and genetic mutations, high- lighting their accuracy and value in complementing imaging data. Combined AI models: a combination of imaging and non‐ imaging Predicting glioma survival with AI by integrating both imaging features with clinical data has further improved prognostic accuracy. For example, adding MRI-based radiomic features to clinical parameters (age, perfor- mance status) reduced the mean absolute error in sur- vival prediction to 4.5 months [40]. A similar study found that radiomics-based AI models further improved pre- dictions by extracting quantitative features from MRI scans, with random forests achieving 92.27% accuracy for PFS. Texture-based features were key to stratifying patients [35]. Another approach achieved 90% accuracy in predicting 1-year survival in glioblastoma patients by combining radiomic and clinical data, demonstrat- ing the value of incorporating patient-specific features for personalised survival estimates. Comparative stud- ies showed that RSFs outperformed other models, with a concordance index of 0.72 for OS prediction. Important features included MGMT promoter methylation and extent of resection [22]. In addition, adding radiomic fea- tures to clinical and genetic data significantly improves survival prediction for low-grade gliomas [41]. Studies have created a nomogram that combines radi- omics signatures from MRI, genetic markers such as IDH mutation, and clinical factors such as age [42]. This combined model showed improved accuracy in predict- ing overall survival compared to models using imaging or non-imaging data alone. In addition, CNNs have been shown to simultaneously process histological images and genomic biomarkers. By exploiting adaptive feedback, the models achieved unprecedented accuracy in predicting glioma survival outcomes, highlighting the potential of multimodal AI frameworks. Such combined approaches are leading the way to precision oncology by providing holistic and individualised survival predictions [43]. The integration of imaging and non-imaging data rep- resents a paradigm shift in survival prediction. A dual graph neural network (GNN) combining radiomic and clinical features has been developed using transformer decoders, achieving a classification accuracy of 0.586 on the BraTS20 dataset [29]. By combining complementary Page 7 of 11Awuah et al. European Journal of Medical Research (2025) 30:61 data modalities, this approach outperforms stand-alone imaging or clinical models. Similarly, ensemble classifi- ers have been used to predict overall survival, IDH muta- tions and other molecular features from a combination of radiomic and clinical data. Here, ensemble methods were shown to consistently outperform individual classifiers [28]. This demonstrates the potential of combined mod- els to improve predictions and guide treatment strategies. This was further illustrated by evaluating the synergy of multimodal data by incorporating tumour location and radiomic features to further improve survival prediction accuracy. The results indicate that combined models not only improve predictive performance, but also provide a more complete picture of the factors influencing glioblas- toma outcomes [44]. Table  3 summarises the strengths and limitations of different parameters used to predict glioma survival. Comparison of the efficacy of imaging, non‑imaging and combined AI methods for glioma survival prediction Imaging‐based models are adept at capturing the fine details of tumour characteristics including shape, texture, and growth patterns and have high predictive accuracy. For instance, post‐radiotherapy MRI‐derived radiomics have been shown to have an AUC of up to 0.93 in pre- dicting glioblastoma survival outcomes [23]. In contrast, non-imaging models are highly accessible and cost effec- tive, using readily available clinical and molecular data such as age and KPS to achieve substantial accuracy with minimal resources [33]. However, whilst imaging mod- els demand advanced technologies and expertise that may not be accessible in all clinical settings, non-imaging models lack the ability to capture critical tumour-specific insights, such as spatial and textural features, which can limit their predictive accuracy. Multimodal imaging has been shown to synergise multiple imaging modalities such as MRI, CT, and PET to achieve accuracies of around 82% [20]. By combining anatomical, functional, and metabolic insights, a compre- hensive understanding of glioma behaviour is achieved. On the other hand, non-imaging models use a variety of data types, including molecular biomarkers and qual- ity of life metrics, to improve patient centric predictions and add substantial value in resource‐limited settings [34]. However, whilst multimodal imaging requires sub- stantial computational infrastructure to integrate these complex data sources, non-imaging approaches are lim- ited by their inability to directly assess tumour physiol- ogy or structural progression, reducing their predictive robustness. Interpretable outputs of DL imaging models, includ- ing 3D CNNs, identify critical brain regions that affect survival, which are essential for personalised treatment strategies [2]. Meanwhile, non-imaging models using patient-reported outcomes, such as quality-of-life data, add a human dimension to survival predictions, address- ing patient-specific concerns often overlooked in imaging models [34]. However, DL imaging models often operate as a “black box” of many algorithms, the hidden layers in neural networks, significantly reducing the interpret- ability of a potentially powerful predictive model, mak- ing its decision-making process opaque to clinicians and limiting trust and widespread adoption [45]. On the other hand, whilst patient-reported outcomes add con- text to survival estimates, they may introduce subjec- tivity and variability that can affect model accuracy and consistency. Finally, combined models that integrate imaging with clinical and molecular data demonstrate the greatest potential by leveraging the strengths of both approaches. As an example, combining MRI‐derived radiomics with clinical data has reduced the mean absolute error in survival predictions to 3.4  months [35]. These models can provide stratification of patients into more precise survival categories so that patients can be treated more personally. Although combined models achieve better accuracy and prediction power, they require large com- putational and logistical resources to integrate and man- age different data types [30]. Non-integrated models, whether imaging-based or non-imaging-based, are easier to implement but lack the nuanced insights necessary for advanced treatment planning. Overall, the most promising approach is the combined method, but practical challenges related to data integra- tion and computational demands need to be overcome for widespread clinical application. Discussion; general limitations and probable solutions for using AI models to predict glioma survival A major challenge of using AI for glioma survival predic- tion is the existence of data heterogeneity which stems from variations in the patients’ characteristics, treat- ments and performance status [43]. This variability limits the model generalisation, particularly when used across different populations. Furthermore, the combination of multiple data types, including histopathological fea- tures and molecular signatures, poses a problem due to the limited sample size and sparsity of the data, increas- ing the risk of overfitting [31, 43, 46]. To overcome this limitation, there is a need to create multi-institutional and common datasets to increase data variability and generalisation. The solution to this problem can be found in federated learning, which allows for model training across institutions whilst keeping patient data private. Page 8 of 11Awuah et al. European Journal of Medical Research (2025) 30:61 Table 3  Parameters contributing to survival prediction in malignant glioma MRI, Magnetic Resonance Imaging; fMRI, Functional Magnetic Resonance Imaging; PET, Positron Emission Tomography; ML, Machine Learning; 3D, 3-Dimensional; DL, Deep Learning; AI, Artificial Intelligence; CNN, Convolutional Neural Networks; CPH, Cox Proportional Hazards; PFS, Progression-free Survival; DNA, Deoxyribonucleic Parameter type Significant contribution Lesser contribution Imaging features [17, 20, 22–26, 30, 32] MRI-based radiomics (e.g. tumour volume, texture) achieving high accuracy (AUC: 0.93) for survival prediction CT imaging with tumour enhancement patterns achieving moder‑ ate C-index (0.74) Multimodal imaging (MRI, CT, PET) achieving ~ 82% accuracy Single-modality imaging (MRI) with lower predictive accuracy (C-index: 0.78) Post-radiotherapy MRI significantly improves prognostic out‑ comes Histopathological imaging with 87% accuracy in specific features Longitudinal MRI data integration further refining survival predic‑ tions Reliance on non-multimodal imaging methods limits systemic insights into tumour biology Preoperative contrast-enhanced MRI radiomics predicting short-term survival with AUC of 0.88 for 3 months and 0.78 for 6 months Multi-channel 3D DL integrating multimodal MRI achieving 90.66% accuracy for survival classification Quantitative imaging features identifying survival categories with up to 98% accuracy Clinical parameters [12, 33, 35, 40] Integration of age, KPS, and performance metrics improving predictions Standalone clinical data often results in reduced predictive accu‑ racy compared to integrated approaches Incorporating quality-of-life metrics reducing mean absolute error in survival predictions to 3.4 months Inconsistent data reporting limits standalone clinical models’ generalisability Demographic features like age and treatment history enhance predictive accuracy Molecular parameters [21, 22, 30, 33, 37] Incorporating biomarkers like IDH1 mutations and CDKN2A alterations enhancing multimodal models Isolated molecular markers without integration yielding inconsist‑ ent results Using MGMT promoter methylation status for stratifying glioblas‑ toma patients Limited utility of single-gene analysis in survival prediction due to tumour heterogeneity Multi-omics integration of mRNA, DNA methylation, and micro‑ RNA profiles improving model adaptability to glioblastoma heterogeneity PET-based molecular features enabling survival stratification Combined approaches [29, 30, 35, 40, 41, 44] Combined imaging and clinical data yielding reduced error in survival predictions (mean absolute error: ~ 3.4 months) Models lacking sufficient integration showing lower reliability in prediction outcomes Addition of radiomic features to clinical and genetic data improv‑ ing prediction for lower-grade gliomas Partial integration approaches with lower multimodal complexity Holistic models combining PET imaging, histopathology, and clinical data for robust survival predictions Ensemble learning combining radiomics, genomics, and clinical features achieving 92.27% accuracy for progression-free survival prediction Incorporation of tumour location into combined models improv‑ ing accuracy in glioblastoma predictions Transformer-based multimodal integration achieving classifica‑ tion accuracy of 0.586 on BraTS20 dataset Combined radiomics and clinical features reduced survival prediction error to 4.5 months AI and ML models [11, 15, 16, 20, 28] 3D CNNs and ensemble regression models excelling in multi‑ modal setups CPH models struggle with high-dimensional data unless comple‑ mented by ML approaches Federated learning enables multi-institutional data use whilst preserving privacy Standalone statistical approaches without ML showing reduced utility for complex datasets Explainable AI methods like SHAP and Grad-CAM improving interpretability Overfitting risks in ML models when applied to small or unbal‑ anced datasets Random survival forests handle mixed data types effectively Neural networks combining structural and functional MRI out‑ performing classical statistical models Ensemble classifiers consistently outperform single models for multimodal glioma survival predictions Page 9 of 11Awuah et al. European Journal of Medical Research (2025) 30:61 Programs like the ReSPOND consortium show how multi-institutional databases can corroborate and extend the applicability of AI algorithms. These efforts indicate the effectiveness of collaborative frameworks in address- ing data heterogeneity and enhancing AI performance in clinical settings [47]. One of the biggest issues with advanced AI models, especially those based on DL methods, is that they are often ‘black boxes’. This lack of transparency and inter- pretability makes it difficult for clinicians to compre- hend and rely on the outcomes yielded by these tools, hampering their clinical implementation [13, 48]. Lack of well-defined decision-making procedures is still a fac- tor that hinders the adoption of AI in practice. Better model interpretability requires using frameworks like SHapley Additive exPlanations (SHAP), Locally Inter- pretable Model-agnostic Explanations (LIME), and Gra- dient-weighted Class Activation Mapping (Grad-CAM) to explain the decision-making process [48, 49]. SHAP employs cooperative game theory to evaluate the con- tribution of each feature to the model’s output. In con- trast, LIME generates locally faithful explanations by modifying data and analysing changes in predictions. Grad-CAM provides visual insights into neural network decision-making, primarily in image classification, by producing heat maps that highlight key regions influenc- ing class predictions. [48, 49]. Furthermore, providing cli- nicians with information regarding how these models are used can help to overcome the gap between technicality and usability. Large-scale prospective studies with stand- ardised protocols are needed to confirm the potential of AI tools in clinical practice. Increasing international part- nerships might also improve external recognition and the use of interpretable AI in various clinical contexts. Whilst models like the RSF and CPH have been shown to be effective in terms of predictive performance, chal- lenges still remain. RSF performs well in handling non- linear interactions but the interpretation of the results is less straightforward when compared to CPH. On the other hand, CPH models are interpretable but may not work as well in complex data processing situations [11]. Recently, there’s a development in the creation of mixed models that would inherit the interpretability of CPH along with the data processing capabilities of RSF for increasing the predictive accuracy. It is such integrated approaches that could provide clinicians with both deci- sion support and accurate prognosis, thereby ensuring the applicability of the approaches in various clinical con- texts [11]. In low-resource settings, there is a major challenge in terms of the available computational power to support the use of AI [50]. Large-scale cloud platforms and high resource requirements of conventional AI models pose accessibility challenges in resource-scarce environments. The recent emergence of light-weight AI models that are designed to work in environments with limited resources can be seen as a solution [49]. Accessibility can also be improved using cloud-based platforms that can scale to deliver the computational resources required for innova- tive AI applications. Such innovations are essential for the implementation of AI in clinical practice across the world to make these technologies available to everyone. Post-training mathematical optimisation can mitigate this issue. Techniques that streamline AI models reduce their memory footprint and latency whilst maintaining accuracy. These optimisations enable the deployment of AI tools on standard consumer-grade CPUs, making them more accessible in resource-limited clinical envi- ronments [51]. Current glioma survival prediction models are plagued by small datasets with low geographical variability, lack of external validation, and absence of large‐scale prospec- tive studies; rendering them non‐generalizable to other clinical applications [52]. To solve these problems, we need to increase international collaborations and data- sets, use federated learning to increase the diversity of the data (whilst maintaining privacy), and use self-super- vised learning to extract useful features from unanno- tated data. For instance, the addition of longitudinal MRI data and clinical variables has increased the accuracy of survival predictions for glioma patients [16, 53, 54]. Moreover, the complexity of large datasets often masks important predictors. Studies have demonstrated that of the 1265 extracted features, only 29 were significant for survival prediction, indicating the necessity of effective feature selection. Recursive feature elimination and cor- relation based selection can be used to reduce data noise, improve predictive power and refine clinical relevance [40]. Finally, as with all AI applications in oncology, the use of patient data to train AI models is a major ethical and legal issue. It is crucial to protect patient privacy when using such information. These issues can be addressed through the use of standard operating procedures for data management and the use of effective anonymisa- tion techniques. In addition, increased clarity in data use agreements promotes trust in the application of AI in healthcare [55]. Acid; RNA, Ribonucleic Acid; PET, Positron Emission Tomography; 3D, Three Dimensional; AUC, Under the Curve; Grad-CAM, Gradient-Weighted Class Activation Mapping; SHAP, Shapley Additive exPlanations; MGMT, O6-methylguanine-DNA Methyltransferase; KPS, Karnofsky Performance Status; IDH, Isocitrate dehydrogenase Table 3  (continued) Page 10 of 11Awuah et al. European Journal of Medical Research (2025) 30:61 Conclusion By integrating complex data sources and providing individualised risk assessments, AI and ML techniques have the potential to significantly improve survival pre- dictions for patients with malignant gliomas. Although AI models predict more accurately than traditional methods, problems of data heterogeneity, model inter- pretability and the need for large, diverse datasets remain. Overcoming these limitations is essential for the clinical adoption of AI-driven tools and provides a pathway to more precise and personalised treatment strategies that may lead to improved patient outcomes in neuro-oncology. Abbreviations AI � Artificial intelligence ML � Machine learning DL � Deep learning MRI � Magnetic resonance imaging CT � Computer tomography 3D � 3-Dimensional OS � Overall survival PFS � Progression-free survival CPH � Cox proportional hazards RSF � Random survival forest C-index � Concordance index SVM � Support vector machines CNN � Convolutional neural network AUC​ � Area under the curve DNA � Deoxyribonucleic acid RNA � Ribonucleic acid PET � Positron emission tomography 3D � Three dimensional SHAP � Shapley Additive exPlanations Grad-CAM � Gradient-Weighted Class Activation Mapping MGMT � O6-methylguanine-DNA Methyltransferase KPS � Karnofsky performance status IDH � Isocitrate dehydrogenase ANN � Artificial neural network Acknowledgements Not applicable. Author contributions Conceptualisation Ideas; W.A.A. Data curation, Writing of initial draft; W.A.A, A.B.J, S.R, P.A.N.B, J.K.T, M.H.S, T.A.R, and O.A. Writing and approval of Final Draft; All authors. Funding None. Data availability No datasets were generated or analysed during the current study. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Received: 7 November 2024 Accepted: 27 January 2025 References 1. Wang Z, et al. Development and validation of a novel DNA methylation- driven gene based molecular classification and predictive model for over‑ all survival and immunotherapy response in patients with glioblastoma: a multiomic analysis. Front Cell Dev Biol. 2020. https://​doi.​org/​10.​3389/​fcell.​ 2020.​576996. 2. Osadebey M, Liu Q, Fuster-Garcia E, et al. Interpreting deep learning models for glioma survival classification using visualization and textual explanations. BMC Med Inform Decis Mak. 2023;23:225. https://​doi.​org/​ 10.​1186/​s12911-​023-​02320-2. 3. Fyllingen EH, Bø LE, Reinertsen I, Jakola AS, Sagberg LM, Berntsen EM, et al. Survival of glioblastoma in relation to tumor location: a statistical tumor atlas of a population-based cohort. Acta Neurochir. 2021;163(7):1895–905. 4. Chua G, Chua K, Chua E, Wong F, Kusumawidjaja G. Tumor location of GBM predicts for survival. Int J Radiat Oncol Biol Phys. 2019;105(1):96–7. 5. Pope W, et al. MR imaging correlates of survival in patients with high- grade gliomas. AJNR Am J Neuroradiol. 2005;26(10):2466–74. 6. Shi W, Wang X, Liu S, Zheng Z, Dong L, Jiang X. A multivariate retrospec‑ tive analysis of high-grade gliomas: survival and prognostic factors. Cancer Med. 2024;13(15): e7456. https://​doi.​org/​10.​1002/​cam4.​7456. 7. Cao MF, et al. A four-gene signature-derived risk score for glioblastoma: prospects for prognostic and response predictive analyses. Cancer Biol Med. 2019;16:595–605. https://​doi.​org/​10.​20892/j.​issn.​2095-​3941.​2018.​ 0277. 8. Ostrom QT, Price M, Neff C, Cioffi G, Waite KA, Kruchko C, Barnholtz-Sloan JS. CBTRUS statistical report: primary brain and other central nervous sys‑ tem tumors diagnosed in the United States in 2015–2019. Neuro Oncol. 2022;24:v1–95. https://​doi.​org/​10.​1093/​neuonc/​noac2​02. 9. Huang Y, Li J, Li M, et al. Application of machine learning in predicting survival outcomes involving real-world data: a scoping review. BMC Med Res Methodol. 2023;23:268. https://​doi.​org/​10.​1186/​s12874-​023-​02078-1. 10. Lamborn K, et al. Progression-free survival: an important end point in evaluating therapy for recurrent high-grade gliomas. Neuro Oncol. 2008;10(2):162–70. https://​doi.​org/​10.​1215/​15228​517-​2007-​062. 11. Qiu X, Gao J, Yang J, Hu W, Lu JJ. A Comparison study of machine learning (random survival forest) and classic statistic (cox proportional hazards) for predicting progression in high-grade glioma after proton and carbon ion radiotherapy. Frontiers in Oncology. 2020. https://​doi.​org/​10.​3389/​fonc.​ 2020.​551420. 12. Zhao R, et al. Machine learning based survival prediction in Glioma using large-scale registry data. Health Inform J. 2022. https://​doi.​org/​10.​1177/​ 14604​58222​11354​27. 13. Poursaeed R, Mohammadzadeh M, Safaei AA. Survival prediction of glioblastoma patients using machine learning and deep learning: a systematic review. BMC Cancer. 2024;24(1):1581. https://​doi.​org/​10.​1186/​ s12885-​024-​13320-4. 14. Gomaa A, Huang Y, Hagag A, Schmitter C, Höfler D, Weissmann T, Brein‑ inger K, Schmidt M, Stritzelberger J, Delev D, Coras R, Dörfler A, Schnell O, Frey B, Gaipl US, Semrau S, Bert C, Hau P, Fietkau R, Putz F. Comprehensive multimodal deep learning survival prediction enabled by a transformer architecture: a multicenter study in glioblastoma. Neurooncol Adv. 2024;6(1):122. https://​doi.​org/​10.​1093/​noajnl/​vdae1​22. 15. Rathore FA, Khan HS, Ali HM, Obayya M, Rasheed S, Hussain L, et al. Sur‑ vival prediction of glioma patients from integrated radiology and pathol‑ ogy images using machine learning ensemble regression methods. Appl Sci. 2022;12(20):10357. 16. Pati S, et al. Federated learning enables big data for rare cancer boundary detection. Nat Commun. 2022. https://​doi.​org/​10.​1038/​ s41467-​022-​33407-5. 17. Jiang Q, Yang X, Deng T, Yan J, Guo F, Mo L, et al. Comprehensive machine learning-based integration develops a novel prognostic model for glio‑ blastoma. Mol Ther Oncol. 2024;32(3):200838. 18. Ghasem Hajianfar A, Atlas Haddadi Avval A, Hosseini SA, Nazari M, Oveisi M, Shiri I, et al. Time-to-event overall survival prediction in glioblastoma https://doi.org/10.3389/fcell.2020.576996 https://doi.org/10.3389/fcell.2020.576996 https://doi.org/10.1186/s12911-023-02320-2 https://doi.org/10.1186/s12911-023-02320-2 https://doi.org/10.1002/cam4.7456 https://doi.org/10.20892/j.issn.2095-3941.2018.0277 https://doi.org/10.20892/j.issn.2095-3941.2018.0277 https://doi.org/10.1093/neuonc/noac202 https://doi.org/10.1186/s12874-023-02078-1 https://doi.org/10.1215/15228517-2007-062 https://doi.org/10.3389/fonc.2020.551420 https://doi.org/10.3389/fonc.2020.551420 https://doi.org/10.1177/14604582221135427 https://doi.org/10.1177/14604582221135427 https://doi.org/10.1186/s12885-024-13320-4 https://doi.org/10.1186/s12885-024-13320-4 https://doi.org/10.1093/noajnl/vdae122 https://doi.org/10.1038/s41467-022-33407-5 https://doi.org/10.1038/s41467-022-33407-5 Page 11 of 11Awuah et al. European Journal of Medical Research (2025) 30:61 multiforme patients using magnetic resonance imaging radiom‑ ics. La Radiol Med. 2023;128(12):1521–34. https://​doi.​org/​10.​1007/​ s11547-​023-​01725-3. 19. Jimenez-Mesa C, Arco JE, Martinez-Murcia FJ, Suckling J, Ramirez J, Gorriz JM. Applications of machine learning and deep learning in SPECT and PET imaging: general overview, challenges and future prospects. Pharmacol Res. 2023;197: 106984. https://​doi.​org/​10.​1016/j.​phrs.​2023.​106984. 20. Luckett PH, Olufawo M, Lamichhane B, Park Ki Yun, Dierker D, Gabriel TV, et al. Predicting survival in glioblastoma with multimodal neuroimaging and machine learning. J Neuro Oncol. 2023;164(2):309–20. 21. Ye L, Gu L, Zheng Z, Zhang X, Xing H, Guo X, et al. An online survival predic‑ tor in glioma patients using machine learning based on WHO CNS5 data. Front Neurol. 2023;14:1179761. 22. Kim Y, Kim KH, Park J, Yoon HI, Sung W. Prognosis prediction for glioblastoma multiforme patients using machine learning approaches: development of the clinically applicable model. Radiother Oncol. 2023;183:109617. 23. Chelliah A, et al. Predicting glioblastoma survival post-radiotherapy using deep learning: a multicenter AI study. Neuro-Oncology. 2024;26(6):1138–51. https://​doi.​org/​10.​1093/​neuonc/​noae0​17. 24. Luo C, Yang J, Liu Z, Jing D. Predicting the recurrence and overall survival of patients with glioma based on histopathological images using deep learn‑ ing. Front Neurol. 2023. https://​doi.​org/​10.​3389/​fneur.​2023.​11009​33. 25. di Noia C, Grist JT, Riemer F, Lyasheva M, Fabozzi M, Castelli M, Lodi R, Tonon C, Rundo L, Zaccagna F. Predicting survival in patients with brain tumors: current state-of-the-art of AI methods applied to MRI. Diagnostics. 2022;12:2125. https://​doi.​org/​10.​3390/​diagn​ostic​s1209​2125. 26. Dilek O, et al. The success of machine learning algorithms developed with radiomic features obtained from preoperative contrast-enhanced MRI in the prediction of short-term survival in patients with glioblastoma. Cukurova Med J. 2021;46(2):706–13. https://​doi.​org/​10.​17826/​cumj.​904688. 27. Zhuge Y, et al. Automated glioma grading on conventional MRI images using deep convolutional neural networks. Med Phys. 2020;47(7):3044–54. https://​doi.​org/​10.​1002/​mp.​14168. 28. Pasquini L, et al. AI and high-grade glioma for diagnosis and outcome pre‑ diction: do all machine learning models perform equally well? Front Oncol. 2021. https://​doi.​org/​10.​3389/​fonc.​2021.​601425. 29. Sun J, Peng Y. The cross-modality survival prediction method of glioblas‑ toma based on dual-graph neural networks. Exp Syst Appl. 2024. https://​doi.​ org/​10.​1016/j.​eswa.​2024.​124394. 30. Babaei Rikan S, Sorayaie Azar A, Naemi A, Bagherzadeh Mohasefi J, Pirnejad H, Wiil UK. Survival prediction of glioblastoma patients using modern deep learning and machine learning techniques. Sci Rep. 2024;14(1):2371. 31. Osman AFI. A multi-parametric mri-based radiomics signature and a practi‑ cal ML model for stratifying glioblastoma patients based on survival toward precision oncology. Front Comput Neurosci. 2019;27(13):58. https://​doi.​org/​ 10.​3389/​fncom.​2019.​00058. 32. Nie D, et al. Multi-channel 3D deep feature learning for survival time prediction of brain tumor patients using multi-modal neuroimages. Sci Rep. 2019;9(1103):1–14. https://​doi.​org/​10.​1038/​s41598-​018-​37387-9. 33. Senders JT, Staples P, Mehrtash A, Cote DJ, Taphoorn MJB, Reardon DA, et al. An online calculator for the prediction of survival in glioblastoma patients using classical statistics and machine learning. Neurosurgery. 2019;86(2):92. 34. Thara T, Thakul O. Comparison of predicted survival curves and personal‑ ized prognosis among Cox regression and machine learning approaches in glioblastoma. J Med Artif Intell. 2023;6:10. 35. Kwiatkowska-Miernik A, Wasilewski PG, Mruk B, Sklinda K, Bujko M, Walecki J. Estimating progression-free survival in patients with primary high-grade glioma using machine learning. J Clin Med. 2024;13(20):6172. 36. Fu J, Singhrao K, Zhong X, Gao Y, Qi SX, Yang Y, et al. An automatic deep learning-based workflow for glioblastoma survival prediction using preoperative multimodal MR images: a feasibility study. Adv Radiat Oncol. 2021;6(5): 100746. 37. Hao Y, Jing X-Y, Sun Q. Cancer survival prediction by learning comprehen‑ sive deep feature representation for multiple types of genetic data. BMC Bioinform. 2023;24(267):1–16. https://​doi.​org/​10.​1186/​s12859-​023-​05392-z. 38. Panesar SS, et al. Machine learning versus logistic regression methods for 2-year mortality prognostication in a small, heterogeneous glioma data‑ base. World Neurosurg. 2019;2:100012. https://​doi.​org/​10.​1016/j.​wnsx.​2019.​ 100012. 39. Sacli-Bilmez B, et al. Identifying Overall Survival in 98 Glioblastomas Using VASARI Features at 3T. Clin Imag. 2023;93:86–92. https://​doi.​org/​10.​1016/j.​ clini​mag.​2022.​10.​011. 40. Rajput S, Kapdi RA, Raval MS, Roy M. Interpretable machine learning model to predict survival days of malignant brain tumor patients. Mach Learn Sci Technol. 2023;4(2):25025–5. https://​doi.​org/​10.​1088/​2632-​2153/​acd5a9. 41. Choi YS, Ahn SS, Chang JH, et al. Machine learning and radiomic phenotyp‑ ing of lower grade gliomas: improving survival prediction. Eur Radiol. 2020;30:3834–42. https://​doi.​org/​10.​1007/​s00330-​020-​06737-5. 42. Tan Y, et al. Improving survival prediction of high-grade glioma via machine learning techniques based on MRI radiomic, genetic and clinical risk factors. Eur J Radiol. 2019;120: 108609. https://​doi.​org/​10.​1016/j.​ejrad.​2019.​07.​010. 43. Mobadersany P, Yousefi S, Amgad M, Gutman DA, Barnholtz-Sloan JS, Velázquez Vega JE, Brat DJ, Cooper LAD. Predicting cancer outcomes from histology and genomics using convolutional networks. Proc Natl Acad Sci USA. 2018;115(13):E2970–9. https://​doi.​org/​10.​1073/​pnas.​17171​39115. 44. Soltani M, et al. Efficacy of location-based features for survival prediction of patients with glioblastoma depending on resection status. Front Oncol. 2021. https://​doi.​org/​10.​3389/​fonc.​2021.​661123. 45. Tewarie IA, Senders JT, Kremer S, et al. Survival prediction of glioblastoma patients—are we there yet? A systematic review of prognostic modeling for glioblastoma and its clinical potential. Neurosurg Rev. 2021;44:2047–57. https://​doi.​org/​10.​1007/​s10143-​020-​01430-z. 46. Jeong J-W, Lee M-H, John F, Robinette NL, Amit-Yousif AJ, Barger GR, Mittal S, Juhász C. Feasibility of multimodal MRI-based deep learning prediction of high amino acid uptake regions and survival in patients with glioblastoma. Front Neurol. 2019;10:1305. https://​doi.​org/​10.​3389/​fneur.​2019.​01305. 47. Davatzikos C, Barnholtz-Sloan JS, Bakas S, Colen R, Mahajan A, Quintero CB, Capellades Font J, Puig J, Jain R, Sloan AE, Badve C, Marcus DS, Seong Choi Y, Lee SK, Chang JH, Poisson LM, Griffith B, Dicker AP, Flanders AE, Booth TC, Rathore S, Akbari H, Sako C, Bilello M, Shukla G, Fathi Kazerooni A, Brem S, Lustig R, Mohan S, Bagley S, Nasrallah M, O’Rourke DM. AI-based prognostic imaging biomarkers for precision neuro-oncology: the ReSPOND consor‑ tium. Neuro Oncol. 2020;22(6):886–8. https://​doi.​org/​10.​1093/​neuonc/​ noaa0​45. 48. Nazir MI, Akter A, Hussen Wadud MA, Uddin MA. Utilizing customized CNN for brain tumor prediction with explainable AI. Heliyon. 2024;10(20): e38997. https://​doi.​org/​10.​1016/j.​heliy​on.​2024.​e38997. 49. Rahman MA, Masum MI, Hasib KM, Mridha MF, Alfarhood S, Safran M, Che D. GliomaCNN: an effective lightweight CNN model in assessment of clas‑ sifying brain tumor from magnetic resonance images using explainable AI. CMES Comp Model Eng Sci. 2024;140(3):2425–48. https://​doi.​org/​10.​32604/​ cmes.​2024.​050760. 50. Zuhair V, Babar A, Ali R, Oduoye MO, Noor Z, Chris K, Okon II, Rehman LU. Exploring the impact of artificial intelligence on global health and enhanc‑ ing healthcare in developing nations. J Prim Care Commun Health. 2024. https://​doi.​org/​10.​1177/​21501​31924​12458​47. 51. Baheti B, Thakur S, Pati S, et al. Optimization of artificial intelligence algo‑ rithms for low-resource/clinical environments: focus on clinically-relevant glioma region delineation. Neuro-Oncology. 2022. https://​doi.​org/​10.​1093/​ neuonc/​noac2​09.​643. 52. Ali MB, Gu IY, Berger MS, Jakola AS. A novel federated deep learn‑ ing scheme for glioma and its subtype classification. Front Neurosci. 2023;23(17):1181703. https://​doi.​org/​10.​3389/​fnins.​2023.​11817​03. 53. Zapaishchykova A, et al. LGG-05. Imaging-based deep learning for event- free survival prediction in patients with pediatric low-grade glioma. Neuro- Oncology. 2023;25:i56. https://​doi.​org/​10.​1093/​neuonc/​noad0​73.​215. 54. Tak D, et al. LGG-15. Deep learning enables longitudinal risk prediction for pediatric low-grade gliomas after surgery. Neuro-Oncology. 2024. https://​ doi.​org/​10.​1093/​neuonc/​noae0​64.​408. 55. Jian A, Liu S, Di Ieva A. Artificial intelligence for survival prediction in brain tumors on neuroimaging. Neurosurgery. 2022. https://​doi.​org/​10.​1227/​neu.​ 00000​00000​001938. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in pub‑ lished maps and institutional affiliations. https://doi.org/10.1007/s11547-023-01725-3 https://doi.org/10.1007/s11547-023-01725-3 https://doi.org/10.1016/j.phrs.2023.106984 https://doi.org/10.1093/neuonc/noae017 https://doi.org/10.3389/fneur.2023.1100933 https://doi.org/10.3390/diagnostics12092125 https://doi.org/10.17826/cumj.904688 https://doi.org/10.1002/mp.14168 https://doi.org/10.3389/fonc.2021.601425 https://doi.org/10.1016/j.eswa.2024.124394 https://doi.org/10.1016/j.eswa.2024.124394 https://doi.org/10.3389/fncom.2019.00058 https://doi.org/10.3389/fncom.2019.00058 https://doi.org/10.1038/s41598-018-37387-9 https://doi.org/10.1186/s12859-023-05392-z https://doi.org/10.1016/j.wnsx.2019.100012 https://doi.org/10.1016/j.wnsx.2019.100012 https://doi.org/10.1016/j.clinimag.2022.10.011 https://doi.org/10.1016/j.clinimag.2022.10.011 https://doi.org/10.1088/2632-2153/acd5a9 https://doi.org/10.1007/s00330-020-06737-5 https://doi.org/10.1016/j.ejrad.2019.07.010 https://doi.org/10.1073/pnas.1717139115 https://doi.org/10.3389/fonc.2021.661123 https://doi.org/10.1007/s10143-020-01430-z https://doi.org/10.3389/fneur.2019.01305 https://doi.org/10.1093/neuonc/noaa045 https://doi.org/10.1093/neuonc/noaa045 https://doi.org/10.1016/j.heliyon.2024.e38997 https://doi.org/10.32604/cmes.2024.050760 https://doi.org/10.32604/cmes.2024.050760 https://doi.org/10.1177/21501319241245847 https://doi.org/10.1093/neuonc/noac209.643 https://doi.org/10.1093/neuonc/noac209.643 https://doi.org/10.3389/fnins.2023.1181703 https://doi.org/10.1093/neuonc/noad073.215 https://doi.org/10.1093/neuonc/noae064.408 https://doi.org/10.1093/neuonc/noae064.408 https://doi.org/10.1227/neu.0000000000001938 https://doi.org/10.1227/neu.0000000000001938 Predicting survival in malignant glioma using artificial intelligence Abstract Introduction Methods Categories of survival prediction models for malignant gliomas ML-based survival prediction models DL models for glioma survival prediction Statistical models Survival predictions and outcomes based on imaging, non‐imaging and combined AI models Imaging AI models Non-imaging AI models Combined AI models: a combination of imaging and non‐imaging Comparison of the efficacy of imaging, non-imaging and combined AI methods for glioma survival prediction Discussion; general limitations and probable solutions for using AI models to predict glioma survival Conclusion Acknowledgements References