Developing Machine Learning Models for Behavioral Coding.

Area of Science:

• Computational Linguistics
• Health Informatics
• Machine Learning

Background:

• Traditional methods for coding patient-provider communication are resource-intensive.
• Accurate behavioral coding is crucial for outcomes-oriented research and clinical practice improvement.

Purpose of the Study:

• To develop a machine learning supervised classification model for automatic coding of clinical encounter transcripts.
• To evaluate the model's efficacy and transferability across different clinical contexts.

Main Methods:

• Eight state-of-the-art machine learning models were evaluated for recognizing communication behaviors using the motivational interviewing framework.
• Data from adolescent weight loss intervention and human immunodeficiency virus (HIV) clinic visits were used for training and testing.
• Semantic and contextual features were incorporated to enhance model accuracy.

Main Results:

• The support vector machine model achieved the highest performance, with F1-scores of .680 in adolescent and .639 in caregiver sessions.
• Incorporating semantic and contextual features improved accuracy to 75.1% (adolescent) and 73.8% (caregiver) of utterances.
• The model demonstrated reliable performance (k = .639) in classifying patient-provider utterances in HIV encounters, comparable to human coders.

Conclusions:

• Validated automatic behavioral coding offers an efficient alternative to traditional methods, accelerating behavioral research.
• Findings can inform clinical practice by providing data-driven communication strategies for clinicians.
• This work is a foundational step towards automated electronic/mobile Health (eHealth/mHealth) behavioral interventions.