Cancer medicine is advancing at an extraordinary pace. New therapies, genomic tests, clinical trials, and treatment guidelines are published constantly, making it humanly impossible for any oncologist to remain current across all cancer types and all available evidence. This creates a widening knowledge gap between what the latest science says is possible and what is actually practiced in day-to-day clinical settings.
The consequences are significant. Patients treated at academic research centers often receive more up-to-date, evidence-based care than those seen at community hospitals, simply because community oncologists lack the time and resources to stay abreast of rapidly evolving literature. This disparity means that many patients may not receive the best possible treatment for their specific cancer.
Traditionally, practicing oncologists have relied on informal peer consultation - calling a colleague who specializes in a particular cancer type - to fill knowledge gaps at critical decision points. But this approach is ad hoc and unscalable: it depends on personal networks, the availability of experts, and the initiative of the treating physician.
Researchers at MD Anderson Cancer Center and IBM Watson Health set out to explore whether artificial intelligence could serve as an always-available, constantly updated expert advisor - providing practicing oncologists with real-time, patient-specific treatment guidance based on the full body of current evidence.
The team designed an AI application called the Oncology Expert Advisor (OEA) to simulate the experience of consulting a cancer subspecialist. Rather than simply searching medical literature, OEA was built around three core functions: dynamic patient summarization, treatment options recommendation, and management advisory.
The Dynamic Patient Summarization module automatically reads a patient's electronic medical record (EMR) and extracts nine key clinical concepts: tumor histology (cell type), resection margins, disease stage, metastasis history, metastatic sites, molecular profile, therapy components received, therapy timeline, and treatment response. Extracting these concepts from unstructured clinical notes required sophisticated natural language processing (NLP), because medical records often contain abbreviations, sentence fragments, and ambiguous phrasing.
The Treatment Options Recommendation module searches OEA's knowledge corpus - built from the full PubMed database of over 23 million published abstracts plus consensus guidelines - to identify which approved therapies and clinical trials are appropriate for the specific patient profile. Each suggestion is linked to the supporting evidence so physicians can evaluate the basis for the recommendation.
The Management Advisory module captures expert best practices for managing patients on specific therapies, addressing questions beyond initial treatment selection such as side effect management, dose modifications, and monitoring protocols. OEA was designed with an iterative agile development cycle, incorporating clinical expert feedback continuously throughout its development.
The algorithms for extracting clinical concepts from patient records were evaluated using F1 scores - a combined measure of precision (fraction of extracted information that is correct) and recall (fraction of correct information that is found). The system achieved F1 scores between 60% and 90% across the nine key clinical concepts after iterative training.
Performance was notably better for non-time-dependent concepts - such as listing which therapies a patient had received - where F1 scores reached 90%-96%. Time-dependent concepts such as therapy history timelines proved more challenging, achieving F1 scores of 63%-65%. The difficulty stems from the lack of explicit time references and imprecise chronological language in clinical notes.
A key design feature was that OEA maintained links to the original source text for every extracted concept, allowing physicians to verify the AI's interpretation and providing a pathway for clinicians to correct errors. This transparency is essential for building clinical trust in AI-assisted tools.
OEA also generated an intuitive longitudinal view of each patient's history, integrating structured data (lab values, mutation results) with concepts extracted from unstructured documents (diagnoses, disease progression notes, toxicity records) into a unified visual timeline. This harmonized view would otherwise require extensive manual chart review by the treating physician.
The first OEA solution was trained specifically for leukemia - including myelodysplastic syndrome, acute myelogenous leukemia (AML), and acute lymphocytic leukemia (ALL). Training data came from approximately 1,000 unique patients treated at MD Anderson's Leukemia Center over the prior two years. The system was trained to recommend the therapy actually prescribed by MD Anderson oncologists as the ground truth.
The treatment recommendation module achieved remarkable performance: 99.9% recall (it identified essentially every appropriate therapy option) with 88% precision for approved therapy suggestions. For clinical trial matching, it achieved 97.9% recall with 96.9% precision against ClinicalTrials.gov listings.
A second OEA solution targeting lung cancer refined the approach by adopting a more nuanced ground truth: rather than treating one treatment per case as correct, clinical experts curated multiple acceptable treatment options for each of 68 clinical cohorts defined by histology, molecular markers, disease stage, and other features. This better reflected the reality that multiple evidence-based treatments may be appropriate for a given patient.
The iterative approach also revealed important lessons: training on historical prescriptions can be misleading because medical knowledge evolves - a therapy considered optimal two years ago may have been superseded. The team recognized that ground truth for AI training in oncology must be expert-curated and current, not simply extracted from historical prescribing patterns.
One of the study's most important contributions is its candid documentation of the challenges encountered in building an AI for cancer care. The team found that clinical notes are far more difficult to process than published literature because they contain private abbreviations, incomplete sentences, and ambiguous references that no algorithm can interpret without deep contextual knowledge.
The complexity of medical concepts also posed challenges. Determining a patient's stage at diagnosis, for example, requires combining tumor size, lymph node status, metastasis information, and cancer-type-specific staging rules - all of which must be extracted from different documents written at different time points. Each logical layer required separate validation against expert knowledge.
The study highlighted the necessity of sustained clinical expert involvement throughout AI development. Clinical experts were needed not just to label training data, but to define the logic underlying each clinical concept, set appropriate time windows for data extraction, and identify failure modes that computational team members would not recognize. AI development in medicine cannot succeed as a purely technical endeavor.
The team also confronted the challenge of rapidly changing medical knowledge: an AI trained on historical data may learn to recommend treatments that were appropriate at training time but have since been superseded by new evidence. Regular retraining and continuous knowledge base updates are essential to maintaining clinical relevance.
The most transformative potential of an AI like OEA lies in democratizing access to expert knowledge. A community oncologist in a rural hospital, without access to subspecialty colleagues or the latest trial data, could use OEA to provide the same quality of evidence-based treatment recommendations that patients at major cancer centers receive. This addresses a fundamental inequity in cancer care.
For leukemia specifically, where treatment algorithms are complex, rapidly evolving, and highly personalized based on genetic markers, an AI advisory tool could be particularly impactful. The combination of cytogenetic analysis, gene mutation profiling, blast counts, and treatment history that defines optimal AML management represents exactly the kind of multi-dimensional reasoning where AI assistance could supplement human judgment most effectively.
The study also demonstrates that OEA can efficiently screen clinical trials - a task that currently requires laborious manual comparison of patient characteristics against complex eligibility criteria. By automating trial matching, an AI advisor could dramatically increase clinical trial enrollment, accelerating the research that ultimately benefits all future cancer patients.
Looking forward, further development of AI advisors like OEA will require solving remaining technical challenges around processing complex temporal data, handling rare cancer types with limited training data, and building physician trust through transparency. But this work demonstrates that the fundamental technical feasibility of AI-augmented cancer care is established.
This study provides the first rigorous technical demonstration that an AI can construct dynamic patient profiles from complex electronic medical records and use them to suggest evidence-based cancer treatments and clinical trials. The results validate the technical feasibility of a concept that has major implications for healthcare equity and quality.
The work is notable for its honest assessment of both successes and limitations. The authors make clear that AI tools like OEA are meant to augment, not replace, clinical judgment. The system is designed to give oncologists better information and more options to consider - reducing the risk that relevant therapies or trials are overlooked - while leaving all actual clinical decisions to the treating physician.
For the field of AI in oncology, this study establishes important benchmarks and lessons learned. Chief among these is that clinical domain expertise must be central throughout AI development, from concept design to algorithm training to performance evaluation. Technical AI talent alone cannot build a reliable clinical tool.
As AI technology matures and as electronic health records accumulate richer datasets, tools like OEA will become more powerful and more broadly applicable. The vision articulated in this paper - of an on-demand expert advisor available to every oncologist for every patient - represents a compelling and achievable near-term goal for artificial intelligence in cancer medicine.