Enhancing neuro-oncology care through equity-driven applications of artificial intelligence

Abstract

The disease course and clinical outcome for brain tumor patients depend not only on the molecular and histological features of the tumor but also on the patient’s demographics and social determinants of health. While current investigations in neuro-oncology have broadly utilized artificial intelligence (AI) to enrich tumor diagnosis and more accurately predict treatment response, postoperative complications, and survival, equity-driven applications of AI have been limited. However, AI applications to advance health equity in the broader medical field have the potential to serve as practical blueprints to address known disparities in neuro-oncologic care. In this consensus review, we will describe current applications of AI in neuro-oncology, postulate viable AI solutions for the most pressing inequities in neuro-oncology based on broader literature, propose a framework for the effective integration of equity into AI-based neuro-oncology research, and close with the limitations of AI.

Artificial Intelligence and Health Equity

Artificial intelligence (AI) leverages computer science and probabilistic reasoning to process data.¹ Within AI’s machine learning and deep learning subdomains, algorithms analyze and “learn” from input data to modify themselves for optimal analytics and predictive performance.¹ Given AI’s unique ability to efficiently process large quantities of heterogeneous data for complex pattern recognition, there has been rapid adoption and widespread application of AI in all medical specialties to aid in efficient diagnosis, treatment, and delivery of care.²

Conversely, numerous studies have identified the potential for AI-based interventions in healthcare to introduce and perpetuate bias thereby exacerbating existing health disparities.^3,4 As such, consideration of the health equity ramifications of AI-based interventions at the outset rather than as a lower-priority afterthought must be applied. To achieve more equitable health and care delivery, attention must be directed towards enhancing the health of those most in need.⁵

The current applications of AI in neuro-oncology have largely neglected to adopt health equity as a guiding principle. Just as training in the scientific method is driven by hypothesis generation and an established analytic plan before data collection, a health equity lens, too, should be applied prospectively and iteratively to AI-based research programs. Furthermore, tracking trends in disparities and progress towards achieving equity using AI requires the field of neuro-oncology to first commit to standardized collection of patient demographic and social determinants of health (SDOH) data as defined by the Center for Medicare and Medicaid Services Framework for Health Equity.⁶ With AI growing exponentially within healthcare, there has never been a more pressing time to consider the relationship between AI and health disparities in neuro-oncology and to develop sustainable guidelines for equitable use.

In this consensus review, we will discuss present applications of AI in neuro-oncology, postulate novel applications of AI to address specific health disparities in neuro-oncology, develop a framework for equitable applications of AI, and explore the limitations of utilizing AI to promote equitable neuro-oncologic care. In doing so, we aim to challenge researchers and clinicians utilizing AI in neuro-oncology studies to incorporate health equity as a necessary pillar.

The Current Applications of AI in Neuro-Oncology

Current applications of AI in neuro-oncology are focused on enhancing tumor diagnosis and more accurately predicting treatment response, postoperative complications, and survival outcomes.

Tumor Diagnosis

Convolutional neural network (CNN)s trained on whole-slide images of tumor-derived tissue samples offer an effective method to diagnose gliomas without relying on resource-limited molecular workups.⁷ Hollon et al. developed DeepGlioma, a diagnostic system that leverages a CNN and stimulated Raman histology to obtain a rapid molecular classification of surgical specimens with high accuracy.⁸

Several central nervous system (CNS) tumors are notorious for having heterogeneous morphology, such as ependymomas and diffuse glioneuronal tumors.⁹ In such cases, DNA methylation may be used to distinguish brain tumor types.¹⁰ Capper et al. trained a machine learning classifier to assign WHO tumor types according to WHO 2021 diagnostic criteria with 92.8% of cases leading to correct classification.¹⁰ These data underscore the promise for future clinical application.^9,10 AI has also shown promise by noninvasively diagnosing brain tumors using radiographic¹¹ and genomic features.¹²

Deriving Disease Prognostications From Radiographic Features:

AI has been employed to prognosticate outcomes for brain tumor patients.⁹ Li et al. developed a radiomics pipeline to identify features most predictive of Ki-67 expression and, in turn, prognosticate patient survival of patients with lower-grade gliomas.¹³ Building on this work, models have been developed to predict overall survival directly, bypassing biomarker correlates.¹⁴ Similarly, artificial neural networks and CNNs based on MRI data have been employed to automate the quantification of tumor burden and characterize patients’ response to treatment.^15,16

Postoperative Complications

AI has demonstrated utility in defining multiparametric relationships driving postoperative impairments following tumor resection. For patients with supratentorial high-grade gliomas, Ghaith et al. employed a random forest model to identify factors associated with 30-day readmission and 30-day reoperation.¹⁷ Furthermore, Fuse et al. used random forest, support vector machine, and light gradient boosting machine models to predict postoperative delayed hyponatremia for patients with pituitary neuroendocrine tumors.¹⁸

Treatment Response

AI is well-positioned to identify molecular subtypes most amenable to available therapies. For a cohort of 488 low-grade glioma patients, Liu et al. used AI to extract cellular morphometric biomarkers from whole-slide images to stratify tumors into 2 subtypes associated with overall survival and treatment response.¹⁹ Specifically, patients with subtype 2 tumors had poor response to treatment. In their seminal paper, Chen et al. created a least absolute shrinkage and selection operator (LASSO) regularized Cox regression model in patients with meningiomas to predict response to radiotherapy.²⁰

AI has also been utilized to predict which patients will benefit most from maximal surgical resection for select brain tumors. Molinaro et al. used recursive partitioning to identify clinical and molecular features associated with overall survival.²¹

While these noteworthy advances demonstrate the power of AI to change clinical neuro-oncology, there has been limited application of AI to address known health disparities in the field to date. If AI is harnessed without an equity lens as a primary pillar, these innovative tools may widen existing disparities. In the following sections, we will discuss novel AI applications in the broader medical landscape which may encourage equity-focused investigations in neuro-oncology.

Health Disparities in Neuro-Oncology and Potential Targeted Applications of AI to Promote Health Equity

Health disparities are rarely isolated from the social and economic structures that render some patient populations more vulnerable to disease than others.²² In neuro-oncology, patient factors, including age, race, ethnicity, sex, sexual orientation, gender identity, weight, and SDOH, contribute to disparities in diagnosis, treatment access, management, and patient outcomes.²³ While closing these widespread differences in outcomes will ultimately require a multifaceted approach that includes addressing prominent SDOH factors,²⁴ AI is a powerful tool that has the potential to make an immediate, positive impact on closing gaps in neuro-oncology care if harnessed responsibly and strategically (Table 1).

Disparities in Molecular Diagnostics

Inequities in diagnosis primarily originate from structural forces, such as the location of hospitals, access to transportation, and insurance status, which dictate if and when patients can receive neuro-oncologic care.⁶⁰ Brain tumor patients belonging to racial and ethnic minority groups, living in lower-income neighborhoods, and insured with Medicaid had more severe disease at diagnosis when compared with majority populations.²⁵ For example, Black patients were twice as likely to require urgent craniotomy compared to White patients (OR: 2.1, P < .0001).²⁵ In addition, Lamba et al. found that patients with glioblastoma who were older had 17% lower odds of MGMT gene promoter methylation testing (>80 vs 60–69, OR: 0.83, P = .04), while privately insured patients had 78% higher odds of testing compared with uninsured patients (OR: 1.78, P < .001).³²

Targeted Applications of AI to Promote Equity in Diagnosis

Many AI investigations have been developed to screen patients and prevent late diagnoses of cancer.²⁶ For example, Mikhael et al. developed a deep learning model to accurately predict a patient’s risk of lung cancer based on a single low-dose CT scan.²⁷ Similarly, AI-powered, smartphone oral cancer screening applications have been used to prevent delayed diagnosis for patients in low-resource areas.²⁹ Even for cancers without screening pathways, AI has shown extraordinary potential to facilitate early diagnosis. Placido et al. created a deep-learning model to predict the risk of pancreatic cancer based on electronic health record data.²⁸ This landmark application of AI paved the way for targeted screening and generates optimism that a similar tool could be built to identify those patients at the highest risk of brain malignancies.

AI could also be utilized to expedite the diagnosis of primary and metastatic brain tumors at each phase of the diagnostic pathway. Cowan et al. developed a patient-facing, machine-learning diagnostic instrument to diagnose migraine with similar accuracy compared to semi-structured interviews with trained headache specialists.³⁰ A similar technology could be tailored to identify patients with high-risk neurological symptoms associated with brain malignancies, prompting early imaging. AI could also aid primary care providers by flagging high-risk features that should warrant urgent referral to a subspecialist.³¹ For patients who require imaging, deep learning models have demonstrated the ability to distinguish between brain metastases and primary brain cancers, which may expedite diagnosis and treatment.⁶¹ Furthermore, machine learning models based on diffusion-weighted imaging can efficiently differentiate pediatric brain tumors.⁶²

Recent applications of AI could aid in addressing disparities in MGMT testing for patients with glioblastoma. Otero et al. developed a random forest-based clinical decision support system to augment histopathology in resource-limited settings and triage patients for advanced molecular testing.³³ This allowed for an accurate diagnosis of primary brain tumors without relying on molecular features.³³ This approach holds promise to speed up timely referral of patients with a likely diagnosis of glioblastoma to specialized neuro-oncology care.

Disparate Access to Standard Treatment

After patients are diagnosed with a brain tumor, disparities exist in access to standard therapy, therapeutic clinical trials, and supportive and palliative care. Older, female, Black, and Hispanic patients were more likely to receive no form of treatment or significantly less than the standard of care.⁴⁰ Similarly, Black patients with glioblastoma were 14% less likely to receive chemotherapy (OR: 0.86, P = .003) and more likely to experience delays from diagnosis to initiation of radiation (treatment delay: ~3 days, P < .0001) and chemotherapy (treatment delay: ~5 days, P < .0001).³⁷ Furthermore, Black patients had significantly higher odds of recommendation against resection of primary brain tumors (meningioma, OR: 1.13, P < .0001, glioblastoma, OR: 1.14, P = .038, vestibular schwannoma, OR: 1.48, P < .0001) independent of potential clinical, demographic, and socioeconomic confounders.³⁴ The substandard quality of care delivered to marginalized patients suggests the potential role of implicit bias and institutionalized discrimination on access to prompt diagnosis and high-quality treatment.³⁵

Socioeconomic forces also drive these disparities. Patients with glioblastoma living in areas with a higher area deprivation index had 58% lower odds of undergoing chemoradiation (OR: 0.42, P = .01).⁴¹ Furthermore, patients with Medicare insurance and older patients were less likely to receive as many radiation fractions or concurrent temozolomide and radiation therapy.⁶³ Lower-income patients faced barriers to definitive treatment for pituitary adenomas.³⁶ Socioeconomic factors also interact with health system-level drivers such that patients in higher-income, lower-poverty areas had better access to high-volume hospitals, defined as performing >50 craniotomies per year.⁶⁴

Targeted Applications of AI to Promote Equitable Standard Treatment

AI may have the ability to reduce disparities in access to surgical care by minimizing the impact that implicit biases play in surgical decision-making, particularly in the setting of clinical uncertainty.³⁵ Prior studies suggest that surgeons overemphasized the risk of comorbidities in Black patients when compared with White patients with the same comorbidities and thus offered glioma resections less frequently.⁶⁵ Standardized clinical decision support systems have already demonstrated the potential to close disparities in treatment provision.⁶⁶ However, these standardized algorithms are limited by manual extraction of data from the electronic health record.³⁵ Thus, through automated data extraction from the electronic medical record, AI models could identify instances in real time where surgeons’ decisions are out of alignment with the standard of care and may be actively subject to bias.³⁵

Frosch et al. developed 4 different machine learning models that predicted the risk of experiencing treatment delays to address disparities in time from cancer diagnosis to treatment initiation.³⁸ A similar approach could identify patients with glioblastoma at the highest risk for treatment delays, prompting the neuro-oncology care team in real-time to prioritize prompt initiation of chemoradiation. However, notably, the LASSO model developed by Frosch et al. performed worse for patients within underrepresented groups and patients living in disadvantaged neighborhoods.³⁸ Adopting equitable model performance as a critical performance metric⁶⁷ and employing bias corrector techniques⁶⁸ can limit disparate model performance to overcome these limitations.

Employing AI to help care teams quickly create radiation and chemotherapy treatment plans for patients with CNS tumors may prevent delays in treatment initiation. Nguyen et al. trained and validated deep learning models to produce radiation treatment plans for patients with prostate cancer quickly in concordance with clinician-formulated plans.³⁹ By complementing physician expertise, such time-saving models could help providers prevent delayed treatment initiation for patients with brain malignancies, including vulnerable patient populations.

Disparities in Clinical Trial Enrollment

For cancers without a proven cure, such as glioblastoma, therapeutic clinical trials offer a pathway for the creation of new therapies. Numerous studies have demonstrated the existence of disparities in cancer clinical trial enrollment based on race, ethnicity, sex, sexual orientation and gender identity, language preference, age, body size, and socioeconomic status.^41–45 Women and NIH-designated minority patients were 15% and 65% under-accrued in glioma clinical trials, respectively, based on incidence-specific enrollment.⁴³ Furthermore, only 20% of glioma clinical trials adequately reported any data on the race of enrolled participants while only 80% reported accrual by sex, which limits a reliable understanding of the extent of the inequality.⁴³ Disparities in access are crucial to address given Melnick’s study demonstrating that for patients with glioblastoma, clinical trial enrollment was an independent predictor of survival regardless of treatment arm.⁶⁹

Targeted Applications of AI to Advance Equity in Clinical Trial Enrollment

Trial Pathfinder is a computational AI framework built to assess patients’ eligibility for cancer clinical trial enrollment and explore the effects of broadening eligibility on survival outcomes.⁴⁶ While recognizing the protective role that clinical trial enrollment offers research participants, Liu et al. determined that relaxing select exclusionary criteria increased the number of eligible patients, particularly more women, and older patients, while maintaining patient safety.⁴⁶ This data illustrates the ability of AI to intelligently simulate the implications of various inclusion–exclusion criteria, including age, BMI, and comorbidity criteria, to maximize equitable enrollment for various underrepresented groups while maintaining patient safeguards, which is especially important given the higher burden of comorbid disease among under-accrued populations.⁷⁰ Furthermore, AI can quantify which patients are underrepresented to guide recruiting efforts.⁴⁷

Despite adequate social support being a known correlate of clinical trial enrollment, implicit biases in clinicians’ subjective assessment of social support can lead to the exclusion of patients from consideration for a trial.⁷¹ Bhatt et al. used natural language processing-aided review of the electronic medical record to assess the extent of social support in patients with advanced cancer.⁴⁸ In the field of social work, AI tools have been implemented to identify patients in need of increased social support, facilitating referrals to social support agencies.⁴⁹

Inequitable Access to Palliative and Supportive Care

Given the high mortality and morbidity of many brain cancers, ensuring adequate access to palliative and supportive therapies is paramount. Several studies illuminate alarming disparities in access to these crucial services. Older, White, female, and lower-income patients with malignant CNS tumors were more likely to enroll in hospice care.⁵³ At the same time, racial and ethnic minorities interfaced with palliative and supportive services less frequently.⁵⁰ The fact that these patients had limited opportunities to receive care focused on optimizing quality of life may contribute to Lamba et al.’s finding that non-White patients with brain metastases received fewer supportive medications to alleviate their symptoms.⁵⁴ For example, Black patients were 19% less likely to receive headache medications (OR: 0.81, P < .001), Hispanic patients were 22% less likely to receive sleep aids (OR: 0.78, P = .01), and Asian patients were 17% less likely to receive anti-emetics (OR: 0.83, P = .004).⁵⁴ The cumulative significance of these disparities is that many patients with brain cancer experience addressable discomfort.

Targeted Applications of AI to Close Disparities in Palliative and Supportive Care

Manz et al. recently determined that a machine learning-based behavioral intervention prompted clinicians to increase the number of outpatient serious illness conversations with cancer patients at high risk of death.⁵² This AI-informed system decreased end-of-life systemic therapy without modifying other outcomes. Similarly, Parikh et al. used gradient boosting and random forest machine learning to analyze electronic health record data for patients with cancer to predict 6-month mortality and identify patients at most significant risk for short-term mortality who would benefit from timely goals-of-care conversations.⁵¹ Similar technology could be employed in neuro-oncology.

To address disparate access to palliative medications for symptom management, neuro-oncology could adopt Raphaeli and Singer’s machine-learning approach for ensuring adequate nutrition for cancer patients.⁵⁵ Raphael and Singer used unsupervised and supervised machine learning methods to develop an individualized decision support system to identify and grade malnutrition in patients with cancer and guide personalized nutritional therapies tailored to the needs of at-risk patients.⁵⁵ An analogous machine learning-based symptom screening tool could help recommend individualized supportive medications for neuro-oncology patients with untreated symptoms.

Disparities in Patient Survival Outcomes

Disparities along the diagnosis-to-treatment pathway ultimately result in inequitable overall survival. Black patients with glioblastoma had 39% higher odds of unplanned readmission within 30 days compared to non-Hispanic White patients (OR: 1.39, P < .001).⁵⁶ Asian and non-Hispanic White patients with glioblastoma had the lowest 90-day mortality compared with other racial and ethnic groups.⁵⁶ Lastly, social and structural determinants of health, such as socioeconomic status, race and ethnicity, insurance status, and access to care, influence disparities in overall survival for patients with brain malignancies.⁵⁸ At 1 year, non-Hispanic White patients had 22% higher risk of mortality compared with Asian and Pacific Islander patients, while non-Hispanic Black patients had 35% greater risk of non-glioblastoma associated death compared with non-Hispanic White patients.⁵⁸

Targeted Applications of AI to Promote Equitable Survival

Improving survival outcomes while closing disparities for patients with brain tumors represents AI’s most exciting potential. Hwang et al. developed a machine learning predictive tool for 30-day readmission among oncology patients that provided novel actionable insights that could be employed by case management or discharge planning teams to prevent readmission.⁵⁷ Notably, while Hwang et al.’s machine-learning model identified a similar distribution of patients at risk for readmission by sex and race, the machine-learning model identified fewer lower-income patients,⁵⁷ exemplifying the type of algorithmic bias that AI researchers must anticipate⁶⁷ and resolve.⁶⁸

Machine learning approaches can identify structural factors contributing to mortality disparities for brain tumor patients and help inspire solutions. In cardio-oncology, Motairek et al. utilized a novel machine learning approach to overcome the limitations of conventional statistical methods by accounting for the intertwining of associations to identify the sociodemographic and environmental factors most influential for cardio-oncology mortality.⁵⁹ This study highlighted the potential for machine learning to inform targeted interventions.

Framework for Equitable Applications of AI in Neuro-oncology

As we postulate how AI can be applied to mitigate health disparities in neuro-oncology, we also propose a 5-part “HEALS” framework, inspired by previously published works,^67,72,73 to guide researchers and clinicians utilizing AI to promote equity in neuro-oncology (Figure 1). This framework is meant to serve as a foundation to build upon as the neuro-oncology field rapidly incorporates AI into patient care.

Health Equity as a Central Pillar of AI Research

• With existing AI research in neuro-oncology largely neglecting AI’s impact on health disparities, current and future AI research must urgently adopt health equity as a central pillar to ensure that AI models are designed to benefit diverse patient populations.

To date, few AI studies have addressed known disparities in neuro-oncology. Furthermore, many published studies lack an evaluation of AI tools’ impact on disparities in care. Therefore, the most pressing and foundational guiding principle is to view combatting health disparities as an essential component of AI-based research, inspired by the first principle of Badal et al.’s published guidelines.⁷² AI innovators should be encouraged to consider the impacts of tool development and use on relevant health disparities within the field. Reviewers for peer-reviewed journals should be empowered to request that AI-based submissions offer commentary and evaluation of how the AI model may promote or detract from gaps in care for vulnerable patient populations. Furthermore, clinicians who use AI innovations should consider whether the application helps to close or widen disparities in patient outcomes.

Evaluate the Impact of all AI Models on Equity

• Neuro-oncology researchers creating new AI models must deliberately evaluate model performance and impact on underrepresented patient populations who may experience inequitable healthcare and outcomes.

After establishing equity as a necessary pillar of all AI-based applications in neuro-oncology, investigators should commit to evaluating the equity impact of AI applications and research on different subpopulations. This requires access to relevant demographic and SDOH variables, which may pose a challenge to some investigators.⁷⁴ Increased efforts to standardize the collection of SDOH variables in the EHR, including staff training on best practices and leveraging patient billing and insurance codes to extract SDOH data, could enhance access to individual-level SDOH data.^6,74 Multiple publicly available databases contain robust SDOH data at the zip code-, census tract-, and county level, which could be combined with patient-specific clinical datasets for model development.⁷⁴ Inclusion of SDOH variables may improve model performance for patients historically excluded in medicine.⁷⁴ If an AI model is originally generated using purely non-demographic clinical data (eg, tumor segmentation imaging dataset), subsequent evaluation of performance in specific patient subpopulations using demographic and SDOH data is critical to elucidate hidden bias that otherwise would go undetected, especially given recent partiality uncovered in imaging-based AI studies.^74,75

Grounded in distributive justice principles, Rajkomar and Hardt et al. propose a valuable framework of “equal outcomes, equal performance, and equal allocation” for evaluating the fairness of AI models.⁶⁷ Ensuring equal outcomes means that all patients benefit equally from an AI model; however, Rajkomar and Hardt et al. contend that model outputs should preferably achieve equalized outcomes, meaning any relevant disparities in outcomes are reduced due to the model.⁶⁷ Equal performance signifies that AI models perform with equal accuracy in underserved populations.⁶⁷ Therefore, merely using the performance metrics of a model, such as sensitivity and specificity, for the whole cohort to justify its strength would be insufficient⁶⁷; yet, published AI applications routinely only report whole-cohort performance metrics. Instead, deliberate calculation of these metrics for underrepresented subpopulations (eg, racial and ethnic minorities, women, LGBTQ+, patients with obesity, patients with lower socioeconomic status, etc.) is required to ensure that model performance does not falter and thus risk widening disparities through ill-informed clinical recommendations.⁶⁷ Finally, equal allocation means that the model ultimately leads to proportionate allocation of resources to patients by need.⁶⁷ Therefore, we encourage neuro-oncology investigators to proactively and iteratively evaluate AI models’ performance, outcomes, and resource allocation for underrepresented patients, particularly to purposefully assess their influence on equity.

Address Data-Driven, Algorithmic, and Human Bias in AI

• Data-driven, algorithmic, and human bias represent distinct sources of bias in AI models, which can negatively impact underrepresented groups and reinforce existing disparities.⁷³

• AI researchers in neuro-oncology can deploy targeted strategies to address each bias type during all phases of model development to prevent bias from translating into flawed research conclusions and disparate care.

Data-driven bias

Data-driven sources of bias in AI include the overrepresentation of privileged patient populations in training datasets due to structural forces that exclude less-resourced groups.⁷³ Training and testing AI technologies on diverse, complete datasets representative of the target population constitute a primary step in addressing data-driven bias.⁷³ As a reflection of structural inequality, existing clinical datasets are primarily composed of male individuals of European descent, while women and racial and ethnic minorities are mainly excluded.⁷⁶ Since AI models often require large amounts of data to discover and learn complex patterns, insufficient data from underrepresented groups leads to biased outputs and relatively low prediction accuracy for these subgroups.⁷⁷ If AI technologies are subsequently used to guide health research and clinical care, inequities in model performance can perpetuate existing health disparities. However, multiple strategies could foster more complete and diverse datasets. For example, clinical data sharing and collaborations across health institutions could be bolstered by data privacy safeguards and external incentives, resulting in more representative datasets and better-performing AI models.⁷⁸ Furthermore, cross-institutional collaborations could facilitate external validation, a critical component of rigorous methodology.⁷⁷ In addition, federated learning could bypass the need for raw data sharing across institutions by instead sharing AI algorithms trained on local data silos which can coalesce synergistically to establish a global model.⁷⁹ Successful implementation of federated learning requires that institutions house local data silos securely, curate them to facilitate effective local model development and sharing, and establish a federated network, which poses barriers for lower-resourced institutions without advanced technical expertise.^80,81 Several groups have developed strategies to democratize access to federated learning, including establishing federated networks at new clinical centers using open-access technologies.⁸¹ These efforts have resulted in the development of resource-limited federated learning paradigms that adapt to local centers’ computational and communication constraints.⁸²

Algorithmic bias

Algorithmic bias directly stems from training on biased datasets, reinforcing patterns entrenched in the inequity.⁷³ In addition to cultivating more diverse datasets to prevent algorithmic bias, Lara et al. argue that bias-correcting techniques can be implemented before, during, or after model development to mitigate disparities in model performance.⁸³

Prior to model development, researchers can employ resampling techniques for underrepresented subgroups to rebalance the dataset to match baseline population proportions across groups.⁸³ For example, Afrose et al. designed a double-priority bias correction technique (DP) to improve prognosis prediction for underrepresented race and age groups while maintaining high performance for majority groups.⁶⁸ Using a model containing all patients, DP iteratively and incrementally replicates samples of underrepresented demographic subgroups to identify a final model with optimal prognosis prediction.⁶⁸

During model development, data augmentation, adversarial training, and transfer learning can be employed to mitigate bias.^83,84 Data augmentation uses generative methods to create synthetic data for underrepresented subpopulations to reduce bias in model performance.⁸³ Adversarial debiasing techniques train, in parallel, a primary model that predicts an outcome of interest and an adversarial model that strives to predict sensitive attributes, such as patient demographics, with the ultimate goal of maximizing predictive performance while minimizing the covert influence of sensitive attributes in the primary model.⁸³ Additionally, transfer learning, the concept that valuable knowledge acquired by a model from solving one task can be transferred to serve as a building block for a new model that solves a related task, has been employed to reduce algorithmic bias.⁸⁴ For example, Gao et al. found that if they trained and tested a prognostic model using an imbalanced Cancer Genome Atlas dataset (80.5% European Americans) or created separate models for each underrepresented ethnic group, the models had lower predictive performance for the data-disadvantaged groups.⁸⁴ However, if they first trained a model on the overrepresented group and transferred acquired knowledge to build unique models for each ethnic group, there was a significant improvement in predictive performance for all groups.⁸⁴

After model development and evaluation of biased predictive performance across groups, post-processing techniques calibrate the models to increase the fairness of predictions among subgroups.⁸³ For example, the equalized odds technique equates false negative and false positive rates for the privileged and non-privileged groups.⁸⁵ The causal analysis technique improves model fairness by removing the causal paths that most contribute quantitatively to biased predictions.⁸⁶

Reducing algorithmic bias for a single-institution AI model does not imply generalizability to other contexts. This underscores the importance of rigorous external validation; otherwise, the applicability of the research findings may not extend beyond a local center’s reach. In neuro-oncology, researchers can evaluate existing and future AI models to assess for bias, increase efforts to diversify training datasets, and employ these bias-correcting techniques to generate more equitably performing AI models that focus on enhancing tumor diagnosis, prognostication, and treatment.

Human bias

Human bias captures how programmers’ and researchers’ biases—influenced by more considerable societal prejudices—infiltrate the conceptualization, development, and usage of AI models.⁷³ Therefore, prioritizing diverse data sets and engineering equitable AI models will only be possible by establishing a culture of equity among teams advancing these technologies. As Norori et al. argue, human bias stems from the same structural forces that govern the diversity, or lack thereof, of model developers and AI research teams, which influence the specific problems AI models are developed to solve.⁷³ Cultivating diversity within and fostering genuine integration of underrepresented scientists into AI research teams,⁸⁷ aided by outreach spanning the educational pathway,²⁴ constitute upstream safeguards against bias. Adapting from the multidisciplinary tumor board model, creating a sustainable infrastructure to facilitate idea sharing and leverage the diverse expertise of AI developers, neuro-oncologists, and health equity research experts could catalyze innovative AI tools that more equitably serve the diverse neuro-oncology patient population.

Link to Larger-Scale Efforts to Address Upstream Drivers of Health Disparities

• In striving to achieve equitable health outcomes in neuro-oncology, AI should be viewed as a tool within a multifaceted approach that targets the underlying forces that contribute to health disparities.

Despite the potential for AI to promote health equity through careful model development, failure to couple applications of AI with efforts to address the structural forces that generate these disparities will prevent AI from manifesting its full potential. Instead, AI should be viewed as a new prong within a robust, multifaceted approach to lessen disparities in neuro-oncology, including patient-, provider-, health systems-, and societal-directed interventions for health promotion.²⁴ Furthermore, AI could be harnessed to project the equity impact of proposed interventions and policies before implementation.⁸⁸

Safeguard Against Ethical Concerns of AI-Based Research

• The field of neuro-oncology must anticipate and address the ethical implications of AI-based research, including concerns about data privacy/handling, reproducibility, and informed consent/patient autonomy.

Many models rely on large datasets, which raises concerns about the risk of data breaches.⁸⁹ Furthermore, AI algorithms may have the capacity to re-identify previously de-identified patient information.⁹⁰ Therefore, drafting clear regulations for safe data handling is critical, such as data anonymization, encryption methods, and secure data-sharing protocols.^89,91 There is increasing concern that incautious applications of AI can lead to irreproducible and clinically non-useful results.⁹² The lack of reproducibility of AI models is largely attributed to leakage of test data into training data, model reliance on irrelevant data artifacts, and poor representation of the diverse population of interest in the test set.^92,93 Reproducibility may be enhanced by better-educating investigators on the fundamentals and limitations of AI technology,⁹⁴ transparent sharing of data and algorithms,⁹² and checklists to uphold rigorous research methodologies.⁹² These safeguards are especially relevant for equitable applications of AI because models focused on underrepresented patient subgroups may be especially prone to overfitting.⁸⁹ Therefore, best practices for model validation include ensuring that the characteristics of the target population, including demographic and SDOH data, are adequately represented in both the training and validation datasets.⁷⁷

Lastly, AI creates new complexities related to informed consent and patient autonomy. “Blackbox” AI algorithms make it challenging to provide patients with sufficient information to achieve informed consent while uneven consent across populations could limit dataset diversity and bias algorithms.⁹⁵ However, if not harnessed responsibly, AI applications could replicate historical violations of informed consent among marginalized populations.⁹⁶ Furthermore, physicians should filter AI-derived clinical recommendations through their expertise and account for patient preference, which is rarely incorporated into AI applications, to provide appropriate patient-centered care while maintaining patient trust.^89,97

Limitations

While we propose creative AI solutions to pressing disparities in neuro-oncology based on previous applications of AI in the broader medical literature, we recognize as a primary limitation of this consensus review that some prior applications highlighted were not explicitly developed to close disparities nor measure success in promoting health equity. Transforming theoretical applications of AI into disparity-closing solutions will require concerted efforts toward equity measure benchmarks.

Furthermore, we recognize that some of these postulated solutions imply the present use or utility of AI, which is not true of most current neuro-oncology practices but may become increasingly relevant in the future. Importantly, we have reviewed heterogeneous AI techniques for promoting health equity in neuro-oncology, including unsupervised and supervised methods. Distinct machine learning methods may be prone to varying degrees of bias⁹⁸; therefore, careful selection of modeling techniques is required. Unsupervised techniques, which identify patterns without preexisting data labeling and structuring, risk the incorporation of inappropriate data in model decision-making that can perpetuate societal biases and widen disparities.⁹⁹ For example, an unsupervised computer vision model replicated racial, ethnic, gender, weight, ableist, and intersectional biases.⁹⁹ In contrast, supervised modeling may permit investigators to label and structure data input and thus retain more control over model output, minimizing the risk of model bias.¹⁰⁰ Techniques have been developed to prevent sensitive attributes, such as race, nationality, and sexual orientation, from influencing AI model decision-making and identify instances when conclusions may be based on flawed data.⁸³

Such methods are especially crucial for equity-driven applications of AI given that bias in AI models can differentially harm underserved populations, from underestimating healthcare needs in Black patients³ to misjudging ICU mortality risk in publicly insured patients.⁴ Therefore, failing to consider equity at each phase of AI-based research is not a neutral oversight but rather risks negligence by contributing to the perpetuation of disparities in care delivery and health outcomes. Therefore, we urge the field of neuro-oncology to adhere to equity-grounded frameworks, such as our proposed HEALS framework or the total product life cycle (TPLC) equity expanded framework,⁷⁷ to overcome the limitations of AI and help lessen disparities.

Conclusion

AI has the potential to be leveraged to combat longstanding disparities in neuro-oncology despite limited applications in the field thus far. By drawing from the broader medical literature to propose creative applications of AI for specific disparities in neuro-oncology, we hope to galvanize the neuro-oncology community to adopt equity as a central pillar of AI-based research and pursue new disparities-focused innovations. In recognizing AI’s dual potential to widen and narrow disparities, we propose a concise, equity-grounded framework to guide researchers and clinicians using AI. This proposed framework can serve as the foundation for the much-needed expansion of equity-driven AI applications in neuro-oncology.

Finally, we call for the establishment of institutional and national committees on equitable applications of AI in neuro-oncology, consisting of diverse experts in the data science, clinical, and health equity fields, to oversee the implementation of critical next steps for the field. Specific next steps include committing to the standard collection of demographic and SDOH variables in neuro-oncology centers and practices; initiating multi-site efforts to increase data diversification for AI model training and testing; investigating and correcting bias in existing and forthcoming AI models; and utilizing AI to strategize and evaluate larger scale efforts that target systemic drivers of inequities in partnership with public health and community collaborators. Through these steps, the neuro-oncology field has the opportunity to harness this novel tool to counter longstanding inequities.

Conflict of Interest

The authors report no conflicts of interest.

Funding

This investigator (S.J.) was supported (in part) by the Division of Intramural Research of the NIH, NINDS. The content is solely the responsibility of the author(s) and does not necessarily represent the official views of the National Institutes of Health.

References