Personalized Ophthalmic Anesthesia: A Regression Analysis of Patient Characteristics, Surgical Profiles, and Anesthesia Protocols for Outcome Prediction

TL;DR

This study uses machine learning to predict recovery time and satisfaction in ophthalmic anesthesia based on patient and surgical data.

Contribution

A high-performing Random Forest model was developed for personalized anesthesia prediction using synthetic clinical data.

Findings

Random Forest achieved an R² of 0.91, MAE of 0.11, and RMSE of 0.14 in predicting outcomes.

Surgical blood loss, BMI, and platelet count were the most important predictors of recovery time.

Synthetic data showed non-normal recovery time distribution, suggesting non-parametric methods are suitable.

Abstract

This study aimed to develop a predictive model for personalized ophthalmic anesthesia by combining patient demographics, surgical profiles, and anesthesia protocols. Exploratory data analysis, inferential statistics, and various machine learning techniques were applied to a synthetic dataset of 350 simulated patient records, each containing 75 clinical features. The primary outcomes included recovery time and satisfaction postoperatively. Correlation matrices, ANOVA F-values, and Recursive Feature Elimination (RFE) were employed for feature selection, with a focus on both clinical relevance and statistical significance. The Random Forest model was found to outperform all other models, achieving an R² of 0.91, MAE of 0.11, and RMSE of 0.14. The most salient predictors of recovery time, identified by SHAP (SHapley Additive exPlanations) analysis, were surgical blood loss, body mass index (BMI), and platelet count. The dataset underwent rigorous preprocessing, including imputation, normalization, and outlier management. The non-normality of recovery time (p < 0.0001) was further confirmed by the Shapiro-Wilk test, suggesting that non-parametric methods would be appropriate. The overestimation of predictive accuracy in the synthetic data, which arises from reduced variability and idealized feature distributions, should be considered preliminary even for high-performing models. Supported the regression-based models’ capability in aiding the personalized anesthesia protocol architecture design. Future projects should incorporate external validation to assess generalizability and clinical utility using external datasets from clinical settings.

Figures