Clinicians Warn Against Relying on AI Without Biological Insight or Clinical Judgment
Artificial intelligence tools are increasingly used to forecast cancer treatment outcomes, but physicians warn that these models fail without biological grounding and clinical context. When used alone, AI often produces unexplained or biased results. Researchers have shown that mathematical models rooted in known immune pathways can offer limited insight into resistance patterns, but only when interpreted through hands-on diagnostics and physician-guided evaluation. In contrast, when used alone, artificial intelligence models produce predictions that cannot be explained or verified in a clinical setting, especially when data is limited or inconsistent. Physicians caution that no model can replace hands-on diagnostics, biological insight, or patient context.
AI Overgeneralizes When Data Is Sparse
Artificial intelligence relies on training from large datasets to make accurate predictions. However, data is often sparse or incomplete in fields like cancer immunotherapy, where patient numbers are limited and outcomes are variable. This makes AI outputs unreliable and difficult to reproduce. Mathematical modeling, also called mechanistic modeling, uses equations representing actual immune signaling processes. These models replicate specific immune events, such as T cell activation, exhaustion, or suppression, making them clinically useful even when datasets are incomplete.
Mechanistic Models Simulate Immune Behavior Without Replacing Clinical Judgment
Using validated equations, mechanistic models simulate biological interactions between immune cells and tumor cells. These tools attempt to predict immune outcomes such as T cell activation, exhaustion, or suppression in response to therapy. In one use case, models trained on a limited melanoma cohort correctly forecasted when immunotherapy resistance would occur in a larger patient group. While these simulations can offer directional insight, they still require clinical interpretation. The usefulness of mechanistic models depends on how well they integrate with physician-guided assessment, not how closely they attempt to mimic biological precision. Their value lies in supporting—not substituting—hands-on, diagnostic reasoning.
Practical Guidelines for Ethical Modeling and Data Sharing
Reliable modeling requires well-curated data, biological clarity, and reproducible code. Even minor data handling errors can undermine predictions. Researchers recommend five safeguards for safe and ethical data use: obtain informed consent, ensure accuracy during data collection, harmonize variables across sources, use public data repositories, and rely on open-source software. These practices support transparent modeling that can be verified across institutions and patient groups.
Bias and Representation in Clinical Modeling
Artificial intelligence models are only as objective as the data on which they are trained. When datasets exclude specific populations, overlook adverse outcomes, or omit unpublished clinical findings, the models reflect those blind spots. Predictions can become skewed, overlooking risk in entire patient groups or reinforcing systemic disparities. This risk increases when scientific publishing, search engine indexing, or funding priorities control what data is visible, accessible, or validated. Mechanistic models reduce this risk by grounding predictions in immune biology rather than inference alone. Still, even the most biologically grounded models can fail when clinical realities are censored, when datasets are skewed, or when entire populations are underrepresented.
Personalized Medicine
Mechanistic models may support a physician’s understanding of immune dynamics, but cannot replace clinical examination or diagnostic judgment. Simulations of T cell activation, suppression, or exhaustion must be interpreted through patient-specific data, history, and physical findings. Personalized treatment planning depends on how well clinicians can align modeled outcomes with what is actually occurring in the patient.
The Doctor-Patient Relationship Cannot Be Replaced
Predictive modeling must serve a therapeutic relationship, not bypass it. As Jeff Clark, ND, explains in his commentary on anecdote-based medicine, effective care depends on more than output from models. It requires time spent with the patient, context-specific interpretation, and clinical intuition grounded in physical examination. Techniques like palpation, auscultation, percussion, and detailed history-taking are indispensable tools that AI and simulations cannot replicate. These practices reveal texture, tone, and nuance in patient presentation central to personalization.
Clinical Storytelling as Diagnostic Input
Narrative medicine builds on this foundation. It trains physicians to interpret a patient’s words, emotions, and metaphors as part of the diagnostic process. As highlighted in NDNR’s discussion of narrative-based care, patient stories often uncover cyclical flares, shifting symptom timelines, environmental exposures, emotional stressors, and undocumented health events that influence immune regulation and treatment outcomes. These insights cannot be captured through data alone. When narrative medicine is integrated with biologically informed modeling, clinicians gain tools that align patient stories with predicted immune responses while still relying on their own diagnostic expertise.
Personalization Requires Human Interpretation
Together, these perspectives affirm that personalized medicine cannot be automated. It requires models that clarify biological dynamics and doctors who interpret them through relationship-based care. When clinicians integrate physical diagnostics with narrative listening and mechanistic modeling, they are best positioned to detect resistance early and guide treatment in a way that honors both immune biology and the human experience.
Further Reading:
– “EBM versus ABM: Anecdotes Do in Fact ‘Add Up’,”
– “Tell Me Your Story: Narrative Medicine in the Therapeutic Relationship,”
– “Biological Pathways in Immunotherapy Resistance,”
References:
1. Bergman DR, Fertig EJ. Virtual cells for predictive immunotherapy. *Nat Biotechnol*. 2025 Apr 14. https://www.nature.com/articles/s41587-025-02583-2
2. Lvovs D, Mahurkar A, White O, Fertig EJ. Ethical data sharing in biomedical research. *Cell Rep Med*. 2025 Apr 15.
