Machine Learning for Future Payment Prediction in Healthcare Revenue Cycle Management

Executive Summary

This research paper focuses on Future Payment Prediction within Healthcare Revenue Cycle Management (RCM). The study addresses a core operational question: can historical, de-identified billing data be used to predict the fraction of an individual healthcare bill that will be paid within 90 days, and can that prediction improve revenue-cycle prioritization?

Broader experiments showed much weaker explanatory power for provider-level prediction (R² = 0.2934) and aggregate revenue forecasting (R² = 0.1246). In contrast, the Future Payment Prediction scenario achieved outstanding results, reaching R² = 0.9191 with a neural network and R² = 0.9102 with a random forest, justifying its focus as the primary operational optimization model.

Choosing the right prediction target

The most important finding in this study is not which model won. It is that two of the three framings of the problem were not worth deploying at all.

The same underlying dataset supports three plausible questions: how much of this specific bill will be paid, how will this provider perform overall, and what will total revenue be next period. Only the first is predictable to a useful standard. Aggregating away from the transaction destroys the very features — payer type, days since billing, prior payment history — that carry the signal.

Abstract

Healthcare revenue cycle teams routinely prioritize accounts receivable using retrospective reports and manual judgment, yet these approaches often fail to identify which patient balances are most likely to remain unpaid. This paper presents a focused industry case study on Future Payment Prediction in healthcare RCM using de-identified historical billing data from a multi-facility healthcare system. The task is formulated as a transaction-level regression problem: predicting the fraction of each bill expected to be paid within 90 days.

Four commonly used machine learning models were compared on a chronological hold-out evaluation: Linear Regression, Decision Tree, Random Forest, and Neural Network. The reported test-set results show strong predictive performance, with Linear Regression achieving MAE = 0.0060243 (R² = 0.7409), Decision Tree achieving MAE = 0.0019225 (R² = 0.8929), Random Forest achieving MAE = 0.0016033 (R² = 0.9102), and Neural Network achieving MAE = 0.0037950 (R² = 0.9191). The findings indicate that nonlinear models substantially outperform linear baselines for transaction-level payment prediction, while Random Forest provides the most attractive balance between absolute error and predictive stability.

Introduction & Operational Significance

Revenue Cycle Management is the financial backbone of healthcare delivery because it connects patient encounters to charge capture, claims processing, reimbursement, and final payment collection. In practice, however, many RCM teams still operate reactively: accounts are followed up after delays emerge, work queues are broad rather than risk-based, and scarce staff time is spent on balances that may have very different probabilities of recovery.

This study argues that predictive analytics can shift RCM from retrospective monitoring to proactive intervention by using historical billing and payment patterns to anticipate future account behavior. Account-level payment prediction is immediately actionable in industry settings because it can support collections prioritization, patient outreach, payment-plan assignment, and near-term cash-flow planning without requiring radical workflow redesign.

Data and Methodological Framework

The study used retrospective financial records from a medium-sized healthcare system spanning 2022–2024. The transaction-level Future Payment Prediction scenario retained over 28,000 usable records after preprocessing. All Protected Health Information (PHI) was removed in line with HIPAA Safe Harbor expectations, which requires the removal of 18 categories of identifiers to protect patient privacy.

The prediction target is the fraction of each bill expected to be paid within 90 days of billing, making the task a transaction-level regression problem on a bounded, normalized target. Features include: total charges, payments made so far, days since billing, payer type, prior payment history, service type, and whether the account was insured or self-pay. To evaluate performance reliably and avoid look-ahead leakage, the data were split chronologically into an 80% training set and a 20% hold-out evaluation set.

# Hyperparameter configurations selected via Grid Search
decision_tree_params = {
    'max_depth': 15,
    'min_samples_split': 2
}

random_forest_params = {
    'n_estimators': 200,
    'max_depth': 20,
    'min_samples_split': 2
}

neural_network_params = {
    'hidden_layer_sizes': (50, 50),
    'activation': 'relu',
    'learning_rate_init': 0.01
}

Results & Performance Patterns

The empirical comparison shows a clear pattern: linear models leave substantial predictive performance unused, while nonlinear models capture the threshold-driven relationships between billing attributes and eventual payment fractions. Random Forest achieved the lowest absolute error (MAE = 0.0016033), suggesting it offers the highest account-level prediction stability. The Neural Network achieved the highest overall fit (R² = 0.9191).

R² and MAE disagree, and the disagreement matters

Ranking by R² and ranking by mean absolute error produce different winners. The neural network explains the most variance, but the random forest is meaningfully more accurate on the average individual account — its MAE is roughly 58% of the network's.

For a collections queue, MAE is the metric that maps onto the job. A specialist works one account at a time, so per-account error is what they experience. R² describes how well the model tracks the overall spread, which is a reporting concern rather than an operational one. This is why the paper recommends Random Forest for deployment despite the network posting the headline number.

Where the model is still weak

A scatter plot of predicted versus actual payments shows that large-balance accounts remain a difficult edge case. Normalized metrics stay strong because the target is a bounded fraction, but the same fractional error on a large balance translates into a much larger raw-dollar deviation.

A production system should therefore never rank on predicted propensity alone. Combining propensity with claim size flags high-value outliers for senior specialist review, which keeps the model useful where it is reliable and keeps humans in the loop where it is not.

Ethical Deployment & Continuous Auditing

In production, the model should be used to trigger supportive financial actions rather than punitive ones. A low predicted payment fraction should activate tailored outreach, payment-plan offers, or financial counseling, not discriminatory billing treatment.

The cleanest deployment path is a nightly batch pipeline that runs inference on open accounts, writes risk scores back into collections queues, and monitors error drift across payers and facilities over time.

Additional project repositories can be found on my GitHub profile (https://github.com/Thanuka9) and professional publications are documented on LinkedIn (https://www.linkedin.com/in/thanuka-ellepola-a559b01aa/).

References

[1] ICTer 2026 Industry R&D Track, official call for papers.

[2] Springer, Instructions for Using the Microsoft Word Proceedings Paper Template.

[3] U.S. Department of Health and Human Services, Guidance Regarding HIPAA Privacy Rule De-identification.

[4] Breiman, L.: Random Forests. Machine Learning 45(1), 5–32 (2001).

[5] Pedregosa, F., et al.: Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011).