A machine learning framework for predicting healthcare insurance claim denials

The denial of healthcare insurance claims imposes a significant burden on providers, leading to decreased revenue and inefficient operations. This research presents a predictive model to assist providers in managing claim submissions.

The framework uses a rule-based claim scrubber that authenticates submissions before sending them, matching them against known denial patterns. It also employs a two-stage supervised pipeline. The study analyses 678,881 professional claim service lines submitted to a single U.S. payer between 2017 and 2024, of which 31.9% were denied. Stage 1 classifies a submitted line as accepted or denied using a Random Forest over 26 features.

On a temporally ordered, claim-grouped held-out partition of 139,615 lines that retains the natural denial prevalence, it attains a precisionrecall AUC of 0.777 (95% CI 0.770-0.784) against a no-skill baseline of 0.330, ROC-AUC 0.865, accuracy 0.794, balanced accuracy 0.784, recall 0.756 and Matthews correlation 0.552. Stage 2 assigns a reason code over a 67-class label space covering 98.8% of denials, reaching 0.269 Top-1 and 0.498 Top-3 accuracy; the same model reaches 0.918 Top-3 when candidates are restricted to the five most frequent codes, and we report both because reason-code accuracy depends heavily on the label space. Because Stage 2 scores only lines Stage 1 has flagged, end-to-end performance is the product of the two stages: the correct code appears among the first three candidates for 38.8% of all denied lines. Class imbalance was addressed with the Synthetic Minority Over-Sampling Technique applied within training folds only; an ablation shows no measurable benefit at this prevalence (difference in average precision -0.0002, 95% CI -0.0009 to +0.0006) and worse calibration, which we report rather than omit. Tuned XGBoost and LightGBM baselines on the identical partition are statistically indistinguishable from the Random Forest.

The solution provides providers with a claim validation and denial-forecasting tool that reduces the denial rate and simplifies the revenue cycle.