CDMS | A Journey Beyond SHAP Explainability in Healthcare Claims Denials

In recent years, the widespread adoption of digital health technologies has transformed healthcare into a data-intensive domain, creating new opportunities for Advanced Analytics and Artificial Intelligence (AI).

Within Claim Denial Management System (CDMS), which involves the workflow of preventing denials, identifying root causes, and applying corrective actions to mitigate their effects, a healthcare claim refers to a formal payment request submitted by a healthcare provider, such as, a hospital, to an insurance company or a third-party payer, for medical care rendered to a patient. Hence, beyond resolving individual denials, effective management is essential to sustain financial stability, reduce inefficiencies, and safeguard patient access to care.

Every year, U.S. private insurers process around 3 billion medical claims. Approximately 15% (450 million) are initially denied, yet 54.3% of those denials are ultimately reversed after multiple appealing cycles. Consequently, hospitals spend an average of US$43.84 contesting each denied claim, amounting to nearly US$19.7 billion annually, of which US$10.6 billion may represent unnecessary administrative effort because the claims should have been paid in the first place (Premier, 2024).

In this scope, the rapid expansion of medical information, transactions, and prescriptions has intensified administrative workloads, often leading to prolonged claim processing, operational inefficiencies, and delayed reimbursement decisions rendered it increasingly difficult and time-consuming for providers to navigate vast datasets and extract actionable insights for informed decision-making. Overcoming these challenges requires adaptability, automation, robust fraud detection, patient-centered strategies, and strict regulatory compliance.

Global trends indicate a growing adoption of AI, Machine Learning (ML), and Advanced Analytics. In healthcare, AI systems represent a promising strategy due to their ability to process large-scale datasets, uncover hidden patterns, generate actionable insights, and support intelligent decision-making, thereby addressing persistent inefficiencies related to resource allocation, administrative demands, and operational performance.

Thus, sophisticated ML models have achieved remarkable predictive performance in predicting claim denials, identify high-risk claims, and support operational decision-making. However, their increasing complexity and black-box nature have significantly reduced their interpretability. As an approach, Explainable Artificial Intelligence (XAI) has emerged to provide transparent justifications and meaningful insights into model predictions, improving their interpretability and supporting more informed decision-making.

Among XAI techniques, SHAP (SHapley Additive exPlanations) is a widely adopted, model-agnostic method for quantifying the contribution of individual features to a model's prediction. Nonetheless, SHAP explains what influenced a prediction rather than why the outcome occurred, it does not distinguish between correlation and causation, limiting its ability to identify the true root causes of an event or to recommend effective corrective actions.

So, How can we identify not only what influenced a prediction, but also what truly caused the denial and what actions could prevent it?

The gap between attribution and action

This may sound abstract, so let's make it more concrete with an example.

As such, XGBoost is a machine learning model that makes predictions by combining many simple decision trees, learning patterns from historical data to infer new cases, allowing it to identify complex patterns in the data with high accuracy.

The figure below shows a SHAP summary plot for an XGBoost model trained to predict claim denials. Features are ordered by their overall contribution to the model's predictions, with each point colored according to the corresponding feature value (red indicates high values, whereas blue indicates low values).

As shown, SHAP demonstrates what the model leans on, namely, higher values in the indicators bureaucracy_time_days, claim_type_oral and claim_total, contribute for pushing the prediction toward denial. Additionally, patient_rejection indicators provide valuable insights into the model's behavior, indicating that patients with a high historical rejection rate are more likely to have their claims denied.

The slowest 20% of claims, defined as those with bureaucracy_time_days exceeding approximately 40 days, comprise 1,996 claims, of which 74.9% are ultimately denied. Claims resolved faster tell the opposite story: those settled within about two-and-a-half weeks are denied only ~10% of the time (9.7% for the fast tier, 10.6% for the typical one), and even claims that drag on for three-to-five weeks stay at 17.3%, still below the 25% overall average. Denial risk barely moves while a claim is processed on time, then rises sharply only once it stalls past the 40-day mark.

Thus, when a slow claim also belongs to a patient with a very high recent rejection history, the two signals compound: all 660 such claims in the data were denied, a 100% denial rate. In other words, the same claim that would sail through in two weeks becomes a near-certain denial once it drifts past six, and a patient who has been rejected before rarely gets the benefit of the doubt.

At first glance, SHAP appears to provide considerable explanation by identifying which features influenced the prediction and quantifying their contribution. But there's an important catch: SHAP explains the model, not the real-world process behind the prediction or the underlying mechanisms that led to the outcome, for instance:

• Is bureaucracy_time_days a cause of denial, or just a symptom that travels with it?

  
  

• If we intervened in reducing the administrative processing time, would the denial risk fall, and by how much?

  
  

To fill in the missing pieces, three complementary perspectives were added on top of the predictive model. Each answers a question that standalone SHAP cannot, and together they provide a complete overview of the contributors behind claim denials.

View	Question it answers	Method
Relational	Which codes carry risk because of the company they keep?	Graph Neural Network on the claim–code graph
Causal	If we intervene on a factor, how much does denial risk actually change?	DoWhy + EconML ForestDRLearner
Counterfactual	If this factor had taken a different value, would this claim still have been denied?	Pearl-style SCM

Relational view: From features to relationships

Revisiting the previous results, the code features landed in that SHAP summary: the implant and drug codes sit near the bottom of the ranking (medication_code_FI0078, FI0077), their dots barely spreading from zero, where individually, they hardly move the score.

As a result, risk emerges from the relationships between codes rather than from individual codes in isolation. Graph Neural Networks (GNNs) are specifically designed to model this type of relational structure.

To make this concrete, let's analyse two drug codes briefly identified by SHAP analysis. On their own, the biologic drug (FI0078) and the chemotherapy drug (FI0077) look unremarkable: each appears on roughly 780 claims and is denied about 47% of the time, only modestly above the 25% baseline. But the 128 claims that carry both codes together are denied 80% of the time (103 of the 128), a 3.2× lift over the base rate. Neither drug is high-risk by itself, the risk lives in their pairing. This is precisely the signal SHAP cannot see, because it scores each feature's contribution independently, and it is exactly what the relational view is built to capture.

In practice, a GNN learns an embedding for each code by message passing, where every code's representation is repeatedly updated from the claims and neighboring codes it co-occurs with. The output is a relational denial propensity: a code's risk in context.

As displayed above, the Top Combinations panel enhance the high-risk procedure pairings, in which hip prosthesis + biologic ranks first, with 86% of claims being denied, followed by knee prosthesis + chemotherapy at 84%. Crucially, each combo carries its top denial reasons ("Charges included in billed procedure," "Incomplete documentation"), what is defined as the beginning of a causal story, not just a correlation.

By checking the Node Detail a single code's risk profile is detailed. Here the biologic drug (FI0078): 784 claims, 47% denied, a 43% GNN relational propensity, its top denial reasons, and the 18 top combinations it participates in:

This illustrates the key advantage of the relational view: the denial risk of a procedure code depends not only on the code itself, but also on the other procedures that accompany it. While the standalone FI0078 presents a 47% denial rate, pairing it with an implant increases that rate to 86%.

The explorer also closes the loop back to SHAP. Drilling into that top combination lists the individual flagged claims behind the 86% rate (a mean predicted denial probability of 81.4% across the 125 scored claims), and expanding any one of them reveals its own per-claim SHAP breakdown. For claim 776709, a rejected claim scored at 96.2%, the dominant drivers are once again bureaucracy_time_days (+1.71) and the patient's recent rejection history (+1.65), the very features that led the global summary. The relational view surfaces the high-risk pairing; SHAP, claim by claim, still explains what tipped each one over.

Overall, the relational view demonstrates that claim denials are not driven solely by individual procedures, but by the interactions between them. By modeling these relationships, GNNs uncover high-risk combinations that would remain hidden when analyzing procedures in isolation, fostering meaningful insights into the patterns underlying claim denials. Even so, it does not explain whether these combinations are the true cause of a claim denial. To answer that question, we need to move beyond patterns and identify the actual causes of claim denials.

Causality: From associated with to changes

Given this, the previous techniques detailed what features and combinations are associated with denials, yet causality remains unanswered.

At this stage, the key question is no longer which factors are associated with denials, but whether intervening on a factor would reduce the likelihood of a denial. The natural next question is interventional:

• If we increased one of factors, how much would the denial rate actually change?

That is the Average Treatment Effect (ATE), an explorer that quantifies to what extent intervening on a factor changes the probability of claim denial. Behind the scenes, it combines DoWhy, EconML, and a placebo refutation to ensure the estimates are robust.

Comparing the SHAP explanations with the causal estimates reveals both agreement and important differences. SHAP ranks bureaucracy_time_days, claim_type_oral, and claim_total among the most influential features driving the model's predictions. The causal analysis confirms that these factors are not only predictive but also have statistically significant effects on the probability of claim denial. In particular, oral claims increase the denial risk by 18.3 percentage points, large claim amounts by 17.8 percentage points, and longer administrative processing times by 4.9 percentage points.

Conversely, in the SHAP summary, features related to pre-authorization and patient rejection history contribute to the model's predictions, suggesting that the model relies on this information, whereas the causal analysis finds no statistically significant effect for their absence after adjusting for confounding factors.

So far, we've learned what the model relies on and which factors genuinely influence denial risk. The final step is making these insights actionable for an individual claim. Instead of asking, "What increases denial risk on average?", we now ask, "Why was this claim denied, and what is the smallest change needed to obtain a different outcome?" Counterfactual analysis addresses exactly that.

Counterfactual root-cause analysis: From changes to actions

The ATE answers the policy question ("what should we generally do?"). The adjuster has a sharper one: "why was this specific denied claim denied, and what's the smallest change that flips it?" That's a counterfactual, and it needs a model of the data-generating process.

The Counterfactual RCA framework builds a graphical causal model with additive-noise structural equations and runs Pearl's abduction–action–prediction: for each denied claim it abducts the exogenous noise, swaps subsets back to "normal," and Shapley-splits the resulting denial anomaly across the upstream drivers.

Across 200 of the denied claims, a clear pattern emerges: administrative processing time is the leading root cause, accounting for 49% of all top attributions, followed by claim amount at 37%.

Notice the reconciliation. The RCA promotes bureaucracy_time_days back to the top, boosting SHAP's per-claim instinct, and it explaining why: for the claims that actually get denied, bureaucracy time and claim amount are the factors whose noise most often tips them over. SHAP told us what the model saw; the RCA tells us which cause is doing the work.

Counterfactual explanations naturally lead to the next step. Once the primary root cause has been identified, this raises a practical question: If we introduce a change in a feature, is it enough to avoid denial?

For instance, if administrative processing time is the dominant root cause, is reducing it actually worth the operational effort? More importantly, would shortening the administrative process translate into fewer claim denials? Algorithmic recourse provides that answer by simulating claim-specific interventions.

The contrast between the two levers is stark. Of the 2,172 denied claims, reducing the administrative processing time alone flips 1,353 (62%) back to approved, the most-reducible dropping from 95.1% → 5.2% (e.g. claim 784,913). The plot below makes the shift concrete: the current denial probabilities (grey) pile up against the ceiling, but once the bureaucracy time is reduced the mass slides left across the decision threshold, 1,353 claims cross into approval, while 819.

That gap is the same story the earlier methods illustrated, where SHAP leaned the "wrong" way on pre-auth (has_pre_auth_False pushing toward approval), and recourse confirms it, where pre-authorization simply isn't on the causal path for most denials. The lever that actually moves the outcome is processing time, and three independent methods agree on it.

Turning Insights into Action: From Claim Prediction to Claim Prevention

Healthcare claim denials are rarely caused by a single factor, due to their intrinsic complexity and elevated interdependence among administrative, clinical, and financial variables. Considering this, and aligning healthcare with novel technologies, the real value of AI in healthcare claims is not predicting that a denial will happen — it is preventing it.

By combining explainability, relational learning, causal inference, and counterfactual reasoning, it enables to identify high-risk claims and understand why they are denied and which operational changes are most likely to prevent them.

In our case, four complementary perspectives converged on the same conclusion: bureaucracy time is not merely a predictive signal, it is the most actionable operational lever for reducing claim denials. This shift from prediction to intervention fosters the possibility to reduce administrative costs, accelerate reimbursements, and ultimately improve both provider efficiency and patient access to care.

As healthcare continues to generate increasingly complex administrative data, the future lies not in building more accurate black-box models, but in developing AI systems capable of transforming predictions into actionable decisions that improve the entire claims management process.