Kan AI diagnostisera vissa sällsynta sjukdomar från elektroniska patientjournaler ?

Diagnostiska följemodeller 2024 fann fall av sällsynta tillstånd som läkare missat både i träningsdata och i levande försök.

Background

Over the past few years several groups have built transformer-based models that read longitudinal EHR sequences and flag patients whose symptom trajectories match curated rare-disease cohorts. In 2023 a system trained on more than 30,000 US patient records achieved a positive predictive value above 0.7 for four lysosomal storage disorders but fell below 0.5 for a rarer glycogenosis subtype, illustrating uneven performance across disorders. A multi-centre study published the same year compared two proprietary LLMs fine-tuned on anonymised records from specialist clinics and found they recovered 79 % of previously missed cases of Niemann-Pick type C while introducing one false positive per ten true positives. Workflows that combine structured billing codes with unstructured clinician notes have shown the biggest gains, yet they remain brittle when applied to centres whose documentation styles diverge from the training corpora. At least one large health-system rollout was paused after an audit revealed clinically significant drift when ICD-10 codes were updated, underscoring the maintenance burden of keeping rare-disease models current.

SOURCE: BMJ, 2024