Clinical study applying machine learning to detect a rare disease: results and lessons learned

Area of Science:

• Medical informatics
• Rare disease diagnostics
• Machine learning applications

Background:

• Machine learning (ML) can enhance patient identification for timely diagnosis and treatment.
• Acute hepatic porphyria (AHP) is a rare disease with available treatments.
• Early detection of rare diseases like AHP remains a challenge.

Purpose of the Study:

• To evaluate a machine learning model's effectiveness in identifying undiagnosed patients with symptoms of acute hepatic porphyria (AHP).
• To assess the feasibility of using ML for rare disease detection in clinical practice.

Main Methods:

• A machine learning model was trained on 205,571 electronic health records from a single center, using 30 known AHP cases.
• The model identified 22 patients exhibiting classic AHP symptoms who had no prior diagnosis or testing.
• Urine porphobilinogen testing was offered to these 22 patients through their clinicians.

Main Results:

• Out of 22 identified patients, 7 agreed to undergo urine porphobilinogen testing.
• None of the 7 tested patients returned positive results for AHP.
• The study highlights a discrepancy between ML-identified potential cases and confirmed diagnoses.

Conclusions:

• The study provides valuable insights into the challenges of using machine learning for rare disease detection, specifically for AHP.
• Further investigation is required to understand the reasons for the negative test results and to improve ML model accuracy.
• Lessons learned will inform future efforts in applying ML to identify patients with AHP and other rare conditions.