Robust Ensemble Learning to Identify Rare Disease Patients from Electronic Health Records.

Area of Science:

• Medical Informatics
• Machine Learning
• Rare Disease Epidemiology

Background:

• Identifying rare disease patients in large electronic health records (EHRs) is crucial but challenging due to limited confirmed diagnoses.
• Existing prediction models struggle with the scarcity of 'gold standard' data for rare conditions.

Purpose of the Study:

• To develop a robust cascade learning methodology for accurate rare disease patient identification from noisy 'silver standard' EHR data.
• To address the challenge of limited confirmed diagnoses in population-scale health databases.

Main Methods:

• A novel cascade learning algorithm combining unsupervised feature selection, supervised ensemble learning, and unsupervised ensemble clustering.
• Utilizing 'silver standard' data (provisionally labeled patients) for model induction.
• Case study on detecting lipodystrophy patients in a national EHR database.

Main Results:

• The proposed algorithm demonstrates superior performance compared to state-of-the-art prediction techniques.
• Successfully identified previously undiagnosed rare disease patients within a large EHR database.
• Validated the efficacy of learning from noisy, unconfirmed diagnostic labels.

Conclusions:

• Cascade learning offers a robust solution for rare disease patient identification in EHRs, even with limited gold-standard data.
• The methodology effectively handles noisy labels, improving prediction accuracy for rare conditions.
• This approach has the potential to significantly enhance rare disease surveillance and diagnosis in large healthcare systems.