[Paper Review] Learning Optimal Personalized Treatment Rules Using Robust Regression Informed K-NN

Ruidi Chen, Ioannis Ch. Paschalidis|arXiv (Cornell University)|Nov 14, 2018

Machine Learning in Healthcare12 references4 citations

TL;DR

This paper proposes a robust, interpretable method for learning optimal personalized treatment rules using Electronic Health Records (EHRs) by combining regularized Least Absolute Deviation (RLAD) regression with K-Nearest Neighbors (K-NN) for outcome prediction. The approach uses RLAD to estimate feature weights for a robust distance metric, enabling stable local nonlinear prediction, and applies a randomized prescriptive policy that improves HbA1c reduction in diabetes and systolic blood pressure control in hypertension beyond standard of care.

ABSTRACT

We develop a prediction-based prescriptive model for learning optimal personalized treatments for patients based on their Electronic Health Records (EHRs). Our approach consists of: (i) predicting future outcomes under each possible therapy using a robustified nonlinear model, and (ii) adopting a randomized prescriptive policy determined by the predicted outcomes. We show theoretical results that guarantee the out-of-sample predictive power of the model, and prove the optimality of the randomized strategy in terms of the expected true future outcome. We apply the proposed methodology to develop optimal therapies for patients with type 2 diabetes or hypertension using EHRs from a major safety-net hospital in New England, and show that our algorithm leads to a larger reduction of the HbA1c, for diabetics, or systolic blood pressure, for patients with hypertension, compared to the alternatives. We demonstrate that our approach outperforms the standard of care under the robustified nonlinear predictive model.

Motivation & Objective

• To address noise, missing data, and nonlinearity in EHRs when learning optimal treatment rules.
• To develop a computationally efficient and interpretable method for personalized medicine that outperforms standard care.
• To ensure robustness against outliers and distributional shifts in clinical data.
• To provide theoretical guarantees on the out-of-sample performance and optimality of the randomized treatment policy.
• To enable application in settings where standard of care is unknown or ambiguous.

Proposed method

• Uses a regularized LAD (RLAD) problem to estimate robust linear coefficients, minimizing absolute deviations and hedging against outliers.
• Constructs a weighted distance metric for K-NN using squared RLAD coefficients as weights, emphasizing features most predictive of outcomes.
• Predicts future outcomes under each treatment by averaging outcomes of the K nearest neighbors in each treatment group.
• Applies a randomized prescriptive policy where treatment probability is inversely proportional to the exponentiated predicted outcome, promoting exploration and robustness.
• Derives a closed-form threshold for treatment recommendation, improving computational efficiency over cross-validation.
• The method is validated on real EHR data from a safety-net hospital for type 2 diabetes and hypertension patients.

Experimental results

Research questions

• RQ1 Can a robust, nonparametric method based on RLAD and K-NN outperform standard predictive and prescriptive models in personalized treatment recommendation?
• RQ2 Does the randomized policy based on predicted outcomes lead to better expected future outcomes than deterministic policies?
• RQ3 How does the proposed method handle noise, missing data, and nonlinearity in real-world EHR data?
• RQ4 Can the method be applied when the standard of care is unknown or variable?
• RQ5 What is the theoretical guarantee on the out-of-sample performance and optimality of the randomized treatment rule?

Key findings

• The RLAD+K-NN model achieved the highest R² (0.52) and lowest MSE (1.34) for predicting HbA1c in diabetic patients, outperforming OLS, LASSO, Huber, and K-NN alone.
• For hypertension, the model achieved an R² of 0.36 and MSE of 159.74, significantly outperforming Huber (R² 0.22) and CART (R² 0.25).
• The method reduced mean absolute error (MAE) by 1% for diabetes and 1% for hypertension compared to the next best method, with median AE improvements of 1% and 2% respectively.
• The randomized policy led to a larger reduction in HbA1c and systolic blood pressure than standard of care under the robust predictive model.
• Feature importance analysis revealed that HbA1c, blood glucose, and systolic blood pressure were top predictors for diabetes, while systolic BP, age, and sodium were key for hypertension.
• The closed-form threshold for treatment recommendation improved computational efficiency over cross-validation-based selection in prior work.