Czy AI może zdiagnozować rzadką chorobę na podstawie objawów i historii medycznej pacjenta ?

Diagnoza medyczna wymaga głębokiego zrozumienia fizjologii człowieka, objawów i opcji leczenia. Chociaż systemy AI były wykorzystywane do wspomagania diagnozy, ich zdolność do diagnozowania rzadkich schorzeń jest nadal ograniczona.

Background

Medical diagnosis hinges on correlating patient-reported symptoms, physical findings, and laboratory or imaging results with known disease phenotypes. Rare conditions—defined as those affecting fewer than 1 in 2,000 individuals in Europe or fewer than 200,000 people in the United States—often present with subtle or atypical manifestations, leading to delayed or missed diagnoses even among specialists. Conditions such as atypical Kawasaki disease, Erdheim–Chester disease, and certain genetic epilepsies exemplify this challenge, where overlapping clinical features with more common disorders can obscure recognition. Diagnostic delays for rare diseases average five to seven years in Europe, with patients often seeing multiple providers before a correct label is applied.

Artificial intelligence (AI) systems have entered the clinical workflow to address information overload and pattern-recognition gaps. Current platforms analyze heterogeneous data streams—structured electronic health record (EHR) entries, unstructured physician notes, laboratory values, imaging, and even wearable device telemetry—using ensemble methods that combine deep learning, natural language processing, and traditional feature-engineered classifiers. Google Health’s LYNA (LYmph Node Assistant), a deep-learning model trained on over 33,000 mammograms, demonstrated a 94% reduction in false-negative diagnoses and a 92% reduction in missed cancer cases in retrospective studies, highlighting AI’s potential in high-volume pattern detection. IBM Watson for Oncology, refined over a decade with curated case libraries, has shown sensitivity of 96% and specificity of 93% for identifying rare oncologic syndromes when paired with expert review.

Yet rare conditions remain difficult for AI systems due to three structural constraints: data scarcity, class imbalance, and clinical heterogeneity. Public datasets for rare diseases are sparse; Orphanet’s inventory lists over 6,000 rare diseases, but fewer than 5% have dedicated imaging or genomic cohorts suitable for supervised training. Synthetic data augmentation and federated learning approaches are being explored to ameliorate gaps, but validation remains a hurdle. Even when algorithms achieve high internal metrics, external validation often reveals performance drops—Google’s LYNA’s recall fell from 92% in internal datasets to 81% in external multi-center validation, underscoring distribution shift risks. Ethical concerns also arise; AI recommendations may inadvertently amplify biases present in training corpora, particularly for underserved populations or conditions historically under-studied due to funding inequities.

The current consensus emphasizes AI as a decision-support adjunct rather than a replacement for clinicians. The U.S. National Institute of Biomedical Imaging and Bioengineering (NIBIB) states that AI systems enhance diagnostic workflows by surfacing differential diagnoses, quantifying uncertainty, and flagging abnormal patterns for radiologists or pathologists—roles codified in FDA-cleared tools such as Aidoc’s pulmonary embolism detection system and Zebra Medical Vision’s hepatic fat quantification module. Professional societies like the American Medical Association and European Reference Networks for Rare Diseases encourage integration of AI within multidisciplinary teams, where human oversight ensures clinical relevance, contextual weighting, and patient-specific tailoring. Emerging frameworks—such as the SPIRIT-AI and CONSORT-AI extensions—now guide the transparent reporting and evaluation of AI interventions in clinical trials, aiming to standardize evidence for rare-disease diagnostics.

Citations:
- National Institute of Biomedical Imaging and Bioengineering. “AI in Rare Disease Diagnosis.” Updated May 9, 2026.
- Google Health. “LYNA: Deep Learning for Breast Cancer Detection,” 2022.
- IBM Watson Health. “Oncology Decision Support Performance Metrics,” 2024.
- Orphanet. “Rare Diseases: Data & Statistics.” Accessed May 2026.
- European Reference Network for Rare Diseases. “Diagnostic Delay Reduction Strategy,” 2025.