AI in Oncology: Predicting Adverse Drug Reactions in Hematologic Malignancies

Adverse drug reactions (ADRs) represent a significant challenge in oncology, particularly in the treatment of hematologic malignancies. These reactions can lead to increased morbidity and higher healthcare costs. The advent of artificial intelligence (AI) technologies such as machine learning, deep learning, and natural language processing (NLP) has opened new avenues for predicting ADRs with greater accuracy and efficiency.

Burden of ADRs in Hematologic Malignancies

ADRs are a significant concern in patients with hematologic malignancies because of the cytotoxicity of anticancer drugs, the high rates of infections, and the complications associated with infections.^1,2 The prevalence of ADRs in patients with hematologic malignancies is high, with up to 85% of patients undergoing treatment experiencing grade 3 or 4 ADRs.³

“For patients with hematologic malignancies, AI systems provide critical real-time monitoring of dangerous complications,” explained Viola Dsouza, PhD, a researcher at Maastricht University specializing in the use of AI for pharmacovigilance. “Algorithms have been developed to predict thrombocytopenia, neutropenia, and cytokine release syndromes based on patient profiles and treatment regimens; for instance, CAR [chimeric antigen receptor] T-cell therapies and kinase inhibitors.”

ADRs can negatively influence treatment outcomes in patients with hematologic malignancies. Blood cancer treatments such as chemotherapy and stem cell transplantation can disrupt the gut microbiota, potentially leading to negative treatment outcomes and increased infection risks.⁴ This microbiota disruption can affect drug metabolism and efficacy, as well as modulate the host immune response, which is crucial for treatment effectiveness.⁴ Furthermore, ADRs impose a financial burden on patients and healthcare systems, increasing healthcare costs due to additional treatments and hospitalizations.⁵

Challenges With conventional ADR Prediction Approaches

Traditional approaches for predicting ADRs rely heavily on clinical expertise and data from pharmacovigilance systems and population-based studies.^6–8 However, these methods often fail to capture the complex interplay of factors contributing to ADR risk in individual patients. The heterogeneity of blood cancers and the variability in patient responses to treatment further complicate the prediction of ADRs using conventional methods.^6–8

“A key barrier to the development of reliable ADR prediction models is data heterogeneity,” noted Dr. Dsouza. “The development of generalizable models in hematologic malignancies is hampered by differences in electronic health record systems, inconsistent terminology, such as ICD versus SNOMED, and inconsistent reporting of lab values or genomic data.”

AI Technologies in ADR Prediction

Bart Westerman, PhD, an Associate Professor at Amsterdam University Medical Center, and his team used data from the FDA Adverse Event Reporting System (FAERS) to train a convolutional neural networks (CNN) model to predict adverse events for combination therapies. According to Dr. Westerman, AI technologies can identify patterns and risk factors based on vast amounts of data from diverse sources and can therefore address some of the challenges with conventional ADR prediction approaches.

What is the role of machine learning in ADR prediction?

Machine learning algorithms, such as logistic regression, decision trees, and artificial neural networks, have been used to predict chemotherapy-induced ADRs using demographic, clinical, and pharmacological data from electronic health records (EHRs).⁹ Jeongah On, of Seoul National University, and colleagues analyzed 6,812 chemotherapy cycles from 935 adult patients treated with four chemotherapy regimens to train three machine learning algorithms (logistic regression, decision tree, and artificial neural networks) to develop predictive models. In validation studies, machine learning models achieved an area under the curve (AUC) of 0.62-0.83 for predicting ADRs, with logistic regression models performing best.⁹

Deep learning, a subset of machine learning that uses artificial neural networks to process and analyze information, has also shown promise in predicting ADRs, owing to its ability to process large and complex datasets.¹⁰ A study using deep neural networks (DNNs) achieved a mean validation accuracy of 89.4% in predicting ADRs across various drugs.¹⁰ The DNN models integrated gene expression data from the Open TG-GATEs database and FDA adverse event reports from the FAERS database.¹⁰

NLP is a deep learning method for extracting and analyzing unstructured data. NLP algorithms pre-trained on clinical notes and discharge summaries have been used in drug safety monitoring.¹¹ A recent study employed a DeBERTa-based NLP model to predict ADR probability from annotated discharge summaries, achieving an AUC of 0.955.¹¹

However, Nicholas Tatonetti, PhD, Associate Professor of Biomedical Informatics at Columbia University Data Science Institute, highlights a key challenge in this area. “The primary challenge in building these models is phenotype extraction from EHRs,” he explained. “Structured fields only tell part of the story; most clinically meaningful detail, particularly about ADRs, is buried in unstructured notes. To address this, we’ve developed LLM-based pipelines and customized algorithms for information extraction.”

Can AI predict ADRs from pharmacovigilance data?

Pharmacovigilance databases are a valuable resource for training data in ADR prediction models, as they contain vast amounts of data on drug safety and adverse events. Machine learning models have been increasingly used to extract patterns and predict ADRs from pharmacovigilance data.¹²

In a recent study, Dr. Dsouza and colleagues synthesized findings from 13 studies using various machine learning algorithms (regression-based, flexible, and ensemble models) to predict ADR in hospitalized patients.¹² Meta-analysis showed pooled sensitivity and specificity of 78.1% and 70.6% for development-only studies, while externally validated models performed better, with 81.5% sensitivity and 79.5% specificity.¹²

However, variations in data quality and heterogeneity in different pharmacovigilance databases could limit the generalizability of machine learning models for ADR prediction.¹² The authors argued that multifactorial models integrating diverse predictors (demographics, lab values, and comorbidities) are needed for improved ADR prediction.¹²

“Recent advances in AI for pharmacovigilance have shown potential in the early detection and prediction of ADRs,” Dr. Dsouza explained. “Machine learning models have been used to retrieve EHRs, unstructured clinical notes, and data from pharmacovigilance databases like VigiBase and FAERS to identify ADR signals that traditional surveillance systems might have missed.”

Can AI predict ADRs for combination therapies?

Predicting ADRs is particularly challenging when patients receive multiple therapies. A team led by Dr. Westerman used CNNs to identify ADR patterns from 15 million patient records.¹³

“Our approach using a CNN model was aimed at capturing subtle higher-order patterns that might occur because of drug combinations,” Dr. Westerman explained. “Overall, drug combinations show an additive effect, so it can be estimated what the frequency of severe adverse events would be with the limitation that this was only assessed in combinations of two drugs.”

The study findings suggest that adverse drug interactions are mostly additive rather than synergistic. Pattern recognition using CNN autoencoders validated that adverse events occur in broad, recognizable patterns rather than as isolated occurrences.¹³ Benchmark analysis confirmed that even drug combinations known to be problematic primarily result in additive effects rather than unexpected toxicities.¹³ The study provides a framework for the use of CNN to evaluate adverse drug interactions in patients receiving combination therapies.

However, Dr. Westerman noted that their approach has limitations. “Our statistical assessment showed that sufficient power is needed to assess the effect of drug combinations because lower frequencies for some combinations led to noisy data,” he explained. “Therefore, the real-world data can be used only for frequently used drug combinations associated with sufficient adverse event data.”

Commenting on their future work, Dr. Westerman said, “We are interested in bringing our approach to a personalized level by incorporating pharmacodynamics and toxicogenomics. This is challenging but also interesting since local drug effects converge to both desired as well as undesired effects in patients.”

Can AI predict ADRs from medication label information?

To address the scarcity of machine-readable resources for ADR information, a team led by Dr. Tatonetti used natural language processing models to extract ADRs from drug labels.¹⁴ The team compiled a machine-readable ADR database termed OnSIDES, which achieved high accuracy (F1 score of 0.90) in extracting ADRs and contains over 3.6 million drug-ADR pairs from 47,211 drug labels.¹⁴  OnSIDES can be used to predict new drug targets, analyze ADRs by drug class, and predict novel adverse events from chemical compound structures.

Dr. Tatonetti described how his team uses AI to advance ADR prediction. “AI enables us to move beyond population averages and find patient-specific drug interactions,” he explained. “We’re using introspective model architectures like sparse autoencoders to help us understand not just what the model predicts but also why, allowing us to identify specific drug-drug and drug-gene interactions associated with adverse outcomes. These insights are being incorporated into the next generation of our widely-used databases, OffSIDES and TwoSIDES v2, which curate and contextualize large-scale evidence of ADRs and interactions from clinical and post-marketing data.”

Integrating Multi-Omic Data for Personalized ADR Prediction

The integration of genomic and other -omic data with clinical information represents a promising frontier in ADR prediction. Commenting on their work on the Molecular Twin project at Cedars-Sinai, Dr. Tatonetti said, “The Molecular Twin project is a flagship initiative where we consent patients [with cancer] to donate biospecimens and clinical data, enabling us to generate multi-omic profiles, including genomics, transcriptomics, and proteomics, at scale.”

The team has enrolled several thousand patients across all cancer types, including hematologic malignancies. Dr. Tatonetti explained that these data feed into predictive models of prognosis and treatment response, and as the dataset matures, it provides a unique opportunity to anticipate ADRs based on a patient’s molecular and clinical context. Dr. Tatonetti added that the combination of depth (omics) and breadth (EHR data) makes this resource a foundation for individualized ADR prediction.

Practical Considerations and Challenges

Although AI models have shown potential in predicting ADRs,^12–14 the implementation of AI tools in clinical practice faces several challenges, including the requirement for extensive model validation.

“Model validation is multifaceted,” explained Dr. Tatonetti. “We emphasize retrospective validation on held-out and external datasets, but we also aim for translational relevance. When possible, we test predictions using prospective laboratory systems, including patient-derived organoids and other model systems. This allows us to assess not only model accuracy but also biological plausibility before considering clinical integration.”

Dr. Dsouza emphasized the importance of regulatory oversight. “Regulatory frameworks are beginning to acknowledge the transformative potential of AI in pharmacovigilance, but oncology presents unique challenges requiring tailored guidance,” she said. “The FDA’s recently proposed framework for AI models used in drug and biological product submissions is a significant first step towards increasing the trustworthiness, transparency, and reliability of AI-driven tools. Likewise, the European Medicines Agency via the Big Data Steering Group has published a multi-year work plan to maximize the use of AI and big data in medicine regulation.”

According to Dr. Dsouza, integrating AI into a hospital setting requires not just technological readiness, but also infrastructure and culture change. “To facilitate smooth integration, hospitals must make investments in decision support systems, a strong IT infrastructure, and workforce training,” Dr. Dsouza said, adding that clinicians must be equipped through targeted education on AI capabilities, limitations, and interpretation to foster confidence and responsible use. Importantly, to maintain clinician trust, the explainability of AI outputs, why a patient is flagged at risk, must be given the highest priority, she added.

Looking ahead, Dr. Tatonetti noted that the future of AI in hematologic malignancies safety lies in integration and personalization. “I anticipate continued growth in multimodal foundation models, better harmonization of real-world data, and increasing use of synthetic data to improve generalizability and fairness,” he said. “But the most exciting shift will be from reactive to proactive safety to identify patients at risk for harm before treatment decisions are made.”

Dr. Dsouza added that international collaboration will be crucial for the successful clinical implementation of AI models to predict ADRs. “International cooperation presents an opportunity to access rare event profiles, diverse patient populations, and drug response that strengthen model training. Collaboration can be facilitated without jeopardizing patient privacy using common data standards, federated learning models, and secure data enclaves. Building high-performing, generalizable AI tools for pharmacovigilance requires pooling data from different nations, especially in rare hematologic conditions,” she concluded.

Westerman and Dsouza report no relevant financial relationships. Tatonetti reports financial relationships with CARI Health (advisor with equity).