# Developing and Validating Machine Learning-Driven Risk Indices to Predict Patient Dropout During Referral, Evaluation, and Waitlisting for Kidney Transplant

## TL;DR

This paper develops machine learning models to predict when patients drop out during kidney transplant processes, aiming to reduce disparities in access.

## Contribution

The study introduces novel machine learning-driven risk indices for predicting patient dropout at multiple stages of kidney transplant care.

## Key findings

- 46% of referred patients did not attend their first transplant evaluation visit.
- ML models achieved AUROC scores of 0.79, 0.71, and 0.76 for dropout prediction at referral, evaluation, and waitlisting stages.
- Socioeconomic and demographic factors were key predictors of dropout risk at each stage.

## Abstract

Transplant is the optimal treatment for kidney failure; however, disparities in access persist. We developed and validated risk indices to predict early dropout at key stages of the transplant-seeking process not captured in national registries.

We included patients referred for kidney transplant at Houston Methodist Hospital between June 2016, and November 2023. We collected demographic, clinical, patient- and contextual-level socioeconomic variables from electronic health records and publicly available census data. We used machine learning (ML) models to predict the characteristics of patients at higher risk of dropping out: (1) at referral (before starting evaluation), (2) in the process of evaluation (before waitlisting), and (3) during waitlisting (before receiving a transplant). Model performance was evaluated using AUROC.

Of 4133 referred patients, 46% did not attend their first transplant evaluation visit. Of 2414 patients who were medically eligible for transplant and started evaluation, 54% did not become waitlisted. Of 2457 waitlisted patients, 31% became inactive on the waitlist. Higher risk patients were consistently older, obese, and socioeconomically disadvantaged, with stage-specific differences: social factors—such as being single, unemployed, less educated, and living in high-deprivation areas—and African American race dominated at referral (AUROC 0.79); clinical comorbidities and both African American and Hispanic ethnicity were prominent at evaluation (AUROC 0.71); and Hispanic ethnicity, smoking, and digital exclusion were key drivers at waitlisting (AUROC 0.76).

ML models effectively identified dropout risk at referral, evaluation, and waitlisting, enabling early identification of at-risk patients. Targeted interventions could reduce disparities, improve evaluation completion, and increase transplant access.