Can AI drug repurposing systems improve patient outcomes in real-world care?

What does the evidence show about AI repurposing improving real-world outcomes?

The strongest evidence comes from large-scale analyses that mimic clinical trials using real-world patient data. A 2021 study developed a deep learning framework that emulated randomized trials on a coronary artery disease cohort of millions of patients, successfully identifying drugs and drug combinations that improved outcomes but were not originally indicated for the disease [5]. This approach directly addresses the question of whether repurposed drugs work in real-world settings by analyzing how patients actually fared. Another major study, published in Nature Communications in 2023, emulated trials for thousands of medications using clinical records from over 170 million patients to find new uses for Alzheimer's disease, highlighting five top-ranked drugs—pantoprazole, gabapentin, atorvastatin, fluticasone, and omeprazole—with potential benefits [8]. These studies show that AI can systematically test many drugs against real-world data, not just generate hypotheses.

For Alzheimer's specifically, a 2024 study used ChatGPT to propose drug repurposing candidates and then tested the top ten in two large clinical datasets (Vanderbilt University Medical Center and the All of Us Research Program). In a meta-analysis, metformin, simvastatin, and losartan were associated with lower Alzheimer's risk [1]. This is a clean example of AI-driven repurposing leading to a measurable outcome—reduced disease risk—in real-world populations. For Crohn's disease, a 2024 study using Danish electronic health records and machine learning identified ten drugs associated with a reduced risk of surgery for intestinal fibrosis, a severe complication, among 9,179 patients followed for up to 24 years [2]. These findings directly link AI-identified candidates to improved clinical outcomes (avoiding surgery) in a real-world care setting.

How do these AI systems actually work to improve patient care?

AI systems improve patient outcomes by efficiently mining two underused sources of evidence: electronic health records and published clinical case reports. A 2026 study introduced TheraMind, a multi-LLM (large language model) system that screened over 10,000 PubMed case reports for non-small cell lung cancer. Using an ensemble of three AI models, it achieved 92% recall and 99.7% specificity in detecting clinically relevant reports—meaning it found almost all the useful cases while almost never flagging irrelevant ones [3][7]. This allows doctors to learn from rare patient experiences that would otherwise be lost in the vast medical literature. Another approach, described in a 2022 review, uses a two-tiered filtering method: first ranking drugs by their predicted binding affinity to disease targets, then matching top drugs with specific real-world scenarios where they might be safe and effective, generating hypotheses that can be tested in trials [11].

A 2025 review on glioblastoma highlights how AI can integrate multi-omics data (genomic, proteomic, transcriptomic) with clinical records to identify biomarkers and repurpose drugs like antipsychotics, antidepressants, and statins for their anti-tumor effects [10]. This is not just theoretical—the review notes that AI-assisted histopathological image analysis reduced diagnostic variation by 30–40% in validation experiments, meaning more consistent and accurate identification of tumor subtypes that could benefit from repurposed drugs [4]. The key takeaway is that AI systems don't replace doctors; they act as powerful filters, surfacing the most promising drug-disease matches from enormous datasets and reducing the noise that makes manual discovery impractical.

What are the limitations and caveats?

Despite promising results, the evidence has important limitations. Most studies are retrospective analyses of existing data, not prospective randomized controlled trials—the gold standard for proving a drug works. A 2025 review on drug repurposing for viral infections notes that while AI has enhanced the identification of novel drug-virus interactions, clinical challenges like drug resistance, dosing optimization, and safety concerns persist, and many computational predictions have not yet been validated in Phase II–IV trials [6]. Similarly, a 2022 review on cancer drug repositioning points out that repurposed drugs often show insufficient single-agent anticancer effects, and pharmaceutical companies are reluctant to invest in the expensive combination studies needed to prove their value [11].

Data quality is another major concern. A 2025 article on optimizing real-world evidence studies emphasizes that electronic health records vary widely in quality, coding standards, and completeness, and that studies using such data are vulnerable to biases like selection bias and confounding [9]. The authors stress that without standardized data and robust methods, AI findings can be misleading. A 2024 study on Alzheimer's drug repurposing using ChatGPT acknowledged that while the AI suggested promising candidates, the findings still require validation in prospective clinical trials [1]. In short, AI repurposing systems are powerful tools for generating hypotheses and identifying signals, but the journey from a signal in real-world data to a proven therapy that improves patient outcomes still requires rigorous clinical testing.

About These Sources

This answer is built on 11 peer-reviewed studies — published from 2021 to 2026, 8 from 2024 or later, 6 in Q1 journals, collectively cited 234 times — selected as the most relevant from 15 studies that passed quality screening, drawn from 63 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions