Transparent deep learning to identify autism spectrum disorders (ASD) in EHR using clinical notes.

Area of Science:

• Artificial Intelligence in Medicine
• Computational Psychiatry
• Machine Learning for Healthcare

Background:

• Machine learning (ML) is widely used for medical diagnosis but often functions as a black box.
• There is a growing need for transparent and clinically aligned diagnostic algorithms.
• Autism spectrum disorder (ASD) is a neurodevelopmental condition with increasing prevalence.

Purpose of the Study:

• To develop a transparent deep learning (ML) approach for diagnosing medical conditions.
• To demonstrate the efficacy of this approach for autism spectrum disorder (ASD) diagnosis.
• To ensure ML outcomes align with clinical diagnostic rules.

Main Methods:

• Utilized unstructured data from CDC surveillance records, labeled with ASD criteria and case labels per DSM-5 rules.
• Compared one rule-based and three deep ML algorithms, along with six ensembles, on a test set of 6773 sentences (35 cases).
• Evaluated criterion and case labeling performance for each ML model and ensemble, comparing case labeling to seven traditional tests.

Main Results:

• The hybrid BiLSTM ML model showed the highest performance for criterion labeling.
• An ensemble of two BiLSTM models achieved the best case labeling, with 100% precision, 83% recall, 100% specificity, 91% accuracy, and 0.91 F-measure.
• The best ML ensemble demonstrated superior overall accuracy compared to existing diagnostic tests.

Conclusions:

• Transparent ML is feasible, even with limited data.
• Deep ML can provide transparent decisions by focusing on intermediate diagnostic steps.
• ML errors at intermediate stages have minimal impact on final outcomes due to data redundancy.