Robust diagnosis recommendation system for Primary Care Telemedicine using long short-term memory multi-class...

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Informatics
• Deep Learning Applications

Background:

• Telemedicine platforms can enhance diagnostic coding support for healthcare providers.
• Efficient diagnosis selection is crucial for timely patient care within virtual consultations.

Purpose of the Study:

• To develop a deep learning-based recommendation system for International Classification of Diseases, 10th Revision (ICD-10) coding.
• To assist telemedicine clinicians in making faster and more accurate diagnosis selections.

Main Methods:

• Utilized a Long Short-Term Memory (LSTM) model for multi-class sequence classification.
• Input features included patient symptoms, complaints, and reasons for consultation requests.
• Trained the model on data from over 2.8 million telemedicine consultations across general medicine, dermatology, and mental health.

Main Results:

• The LSTM recommender achieved an average accuracy of 31.7% across general medicine, dermatology, and mental health specialties.
• Demonstrated an average coverage of 85.8% within the top 20 recommended diagnoses.
• Achieved an average personalization score of 0.87, indicating tailored recommendations.

Conclusions:

• LSTM multi-class sequence classification effectively recommends diagnoses tailored to individual telemedicine consultations.
• The system is retrainable and can reduce the time and resources providers spend searching for diagnosis codes.
• The recommender demonstrates robustness across diverse clinical specialties, enhancing diagnostic workflows.