AI Flags Insurance Claim Denials Before They Happen With Surprising Accuracy

Every denied health insurance claim sets off a costly chain reaction: administrative reprocessing costs climb by roughly 15 percent per case, reimbursement timelines stretch past 30 days, and tensions between providers and payers deepen. Now, a team of researchers has built a machine learning framework that can predict whether an insurance claim will be denied before it is ever submitted, and even tell billing staff why. The study, published in Frontiers in Health Services, analyzed 678,881 professional claim service lines sent to a single U.S. payer between October 2017 and July 2024, of which 31.9 percent were denied, and demonstrated a two-stage pipeline that turns that historical record into a practical, pre-submission warning system.

The framework, called PF-CDM, works in two stages. The first is a binary classifier built on a Random Forest, a type of ensemble model that trains hundreds of decision trees on random subsets of the data and averages their votes. It takes in 26 features engineered from 32 raw claim fields spanning temporal, clinical, demographic, provider and financial dimensions, and outputs a probability that a given service line will be denied. The second stage is a separate multiclass Random Forest trained only on denied lines, which assigns one of 67 denial reason codes covering 98.8 percent of the denials in the dataset. Before either model runs, a rule-based scrubber validates the submission, checking for missing procedure codes, invalid dates and other errors that guarantee rejection.

The performance numbers are striking in context. On a held-out set of 139,615 lines, deliberately kept at the natural denial prevalence of 33 percent and ordered in time rather than shuffled, the first-stage model achieved a precision-recall AUC of 0.777, compared with a no-skill baseline of just 0.330. Its ROC-AUC stood at 0.865, accuracy at 0.794 and Matthews correlation coefficient at 0.552. At the chosen operating threshold of 0.469, the model correctly caught 34,850 denials while missing 11,250 and raising 17,574 false alarms. The team used percentile bootstrap confidence intervals over 1,000 replicates resampled by claim rather than by line, since lines of one claim are adjudicated together and are not statistically independent.

Perhaps the most refreshingly honest result concerns a technique most practitioners treat as gospel. The researchers applied SMOTE, the Synthetic Minority Over-sampling Technique, which creates synthetic minority-class examples by interpolating between existing ones, but only inside training folds so that no synthetic record was ever scored. Then they ablated it. At a denial prevalence of 31.9 percent, the difference in average precision between the resampled and unsampled models was just 0.0002, with a 95 percent confidence interval spanning zero, and the unsampled model was actually better calibrated, with a Brier score of 0.140 versus 0.145. The authors retained SMOTE in the reported pipeline but state plainly that the evidence shows it is unnecessary at this prevalence, a finding that could save other teams considerable computational effort.

Equally rigorous was the assault on data leakage, the silent killer of machine learning projects. The team partitioned the data temporally, with all claims created before 21 July 2023 forming the development set and everything after forming the test set, and grouped on the claim so that lines of the same claim never straddled the split. They removed the denial amount from the feature space because it was populated only for denied lines and would have reproduced the target almost exactly. They also cut off historical aggregate features at fixed intervals before each index claim: the provider history signal stabilized at 90 days, while the patient-history curve was still declining at 365 days, an effect the authors report as an upper bound rather than an estimate. A shuffled-label control, in which a model trained on randomly permuted labels scored a mean ROC-AUC of 0.4835, straddling chance, confirmed that no feature encoded the outcome of its own line.

The second stage tells the more sobering story. Predicting the reason code over the full 67-class label space is materially harder than the five-code formulations common in earlier literature, and the model achieved 0.269 Top-1 accuracy and 0.498 Top-3 accuracy, against a majority-class floor of 0.142. Restricted to the five most frequent codes, the same model reached 0.918 Top-3 accuracy, which is precisely why the authors insist that reason-code results are not comparable across studies unless the label space is stated. Because stage two only scores lines the first stage has flagged, the end-to-end cascade is necessarily weaker than either stage alone: the correct reason code appears among the top three candidates for 38.8 percent of all denied lines. Crucially, the errors are systematic rather than random, with the model confusing codes that are adjacent in adjudication practice, such as the three coverage and eligibility codes it mixes up with one another.

The feature analysis delivers an insight that will surprise billing veterans: claim economics, not provider identity, carry the signal. When entire feature categories were withheld, dropping the four financial features cost 61.5 percent of the total predictive loss, while removing provider features changed precision-recall AUC by exactly nothing. SHAP attributions, computed on a constrained surrogate model whose fidelity was measured against the deployed forest, ranked the total charge of the claim, the submission lag, the line’s share of the claim total and the number of lines on the claim as the strongest contributors. In other words, denial risk in this dataset is driven by what is billed and how a line sits within its claim, not by who submitted it.

The framework’s practical value is best expressed in workload terms. Flagging the 20 percent of lines with the highest predicted risk would surface 23,189 of the held-out period’s denials at the cost of 4,734 unnecessary reviews, roughly one false alarm for every five denials caught. Flagging 30 percent raises the yield to 30,565 denials but triples the false alarms. The choice between these operating points, the authors note, is a staffing decision rather than a modeling one, and the model’s contribution is that it makes the trade-off explicit and tunable. A worked example shows how a biller might use this: a line the model wrongly flagged with a denial probability of 0.62 was attributed to its low share of a large claim, a pattern associated with bundling denials; a reviewer would simply confirm the bundling status and release it.

The team is equally candid about the limits. All data come from a single U.S. payer concentrated among 123 rendering providers, mostly in Illinois and Michigan, so the labels encode that payer’s policies and conventions. Under a provider-held-out split, precision-recall AUC falls from 0.777 to 0.677, the realistic expectation for an unfamiliar practice. The study covers professional CMS-1500 lines only, and the extract carries neither race nor ethnicity fields, so the disparities most often examined in the fairness literature simply cannot be measured, a gap the authors call the most significant in the evaluation. Tuned XGBoost and LightGBM baselines proved statistically indistinguishable from the Random Forest on identical partitions, meaning the real contribution lies in the architecture, the scrubber and the evaluation protocol rather than the choice of learner.

What makes this study worth watching is its methodological honesty as much as its results. The authors show that a model trained to predict one payer’s decisions will reproduce any bias in those decisions, that a real denial often carries several reason codes while the model assigns one, and that a simple eligibility check before submission could prevent the second most frequent denial outright, hospice enrolment, which accounted for 7.95 percent of denials and requires no model at all. The path to deployment, they argue, runs through a prospective trial proving that presenting these predictions to billing staff actually reduces denial rates, plus the coverage and eligibility validation that the error analysis identifies as the largest available gain. For an industry hemorrhaging revenue through avoidable rejections, a calibrated, explainable early-warning system, even a modest one, may prove transformative.

Subject of Research: Machine learning prediction of healthcare insurance claim denials and denial reason codes

Article Title: A machine learning framework for predicting healthcare insurance claim denials

Article References: Jadoon, M. M., Miakhil, O., Chakari, N. R., Qureshi, M. B., Ullah, I., Ahmed Haider, S., & Ur Rehman, M. (2026). A machine learning framework for predicting healthcare insurance claim denials. Frontiers in Health Services, 6, Article 1927175. https://doi.org/10.3389/frhs.2026.1927175

Image Credits: AI Generated

DOI: 10.3389/frhs.2026.1927175

Keywords: machine learning, health insurance, claim denials, random forest, SMOTE, SHAP, revenue cycle management, predictive modeling, healthcare finance, explainable AI, data leakage, class imbalance