Healthcare cost prediction: Leveraging fine-grain temporal patterns

Area of Science:

• Health Informatics
• Data Science
• Machine Learning

Background:

• Accurate healthcare cost prediction is crucial for financial planning and resource allocation.
• Traditional methods often use coarse-grain data, potentially missing valuable temporal insights.

Purpose of the Study:

• To develop and evaluate a novel method for predicting individual healthcare costs using fine-grain temporal data.
• To assess the impact of spike detection features on prediction accuracy.
• To determine the contribution of cost, visit, and medical data to prediction performance.

Main Methods:

• Utilized three years of healthcare claims data (2013-2016).
• Extracted fine-grain temporal features from cost, visit, and medical information.
• Developed novel spike detection features to capture temporal patterns.
• Applied Gradient Boosting for prediction and compared against baseline methods using Mean Absolute Percentage Error (MAPE).

Main Results:

• Fine-grain predictors achieved a significantly lower MAPE (3.02) than coarse-grain predictors (8.14).
• Incorporating spike detection features further reduced MAPE to 2.04.
• Cost and visit data were significant predictors, while medical data showed minimal impact.

Conclusions:

• Fine-grain temporal patterns substantially improve healthcare cost prediction accuracy.
• Spike detection features enhance predictive performance by capturing temporal dynamics.
• Gradient Boosting is a highly effective model for this prediction task.