Exploring supervised machine learning approaches to predicting Veterans Health Administration chiropractic service...

Area of Science:

• Health Services Research
• Machine Learning in Healthcare
• Spinal Pain Management

Background:

• Chronic spinal pain affects millions of US adults, incurring significant healthcare costs.
• Conservative interventions like chiropractic care may reduce costs and improve pain status.
• Predicting healthcare service utilization for spinal pain, particularly chiropractic use, is not well understood.

Purpose of the Study:

• To explore supervised machine learning for predicting one-year chiropractic service utilization in veterans.
• To assess the feasibility of predicting the frequency of chiropractic care use.

Main Methods:

• Retrospective cohort study of 19,946 veterans using VA chiropractic services.
• Primary outcome: one-year chiropractic service utilization (quartiles: 1, 2-3, 4-6, 7+ visits).
• Compared four machine learning algorithms (gradient boosted, SGD, SVM, ANN) using 158 features.

Main Results:

• Algorithms demonstrated poor prediction capabilities, with subset accuracy ranging from 38.6% to 42.1%.
• Area under the precision-recall curve was modest (0.38-0.43).
• Models showed only a small improvement (approx. 15%) in prediction probability over chance.

Conclusions:

• Supervised machine learning currently has limited clinical utility for predicting chiropractic service utilization.
• Predicting healthcare service utilization for conservative spinal pain interventions remains challenging.
• Future research should focus on enhancing model performance for better clinical application.