State-of-the-art for automated machine learning predicts outcomes in poor-grade aneurysmal subarachnoid hemorrhage using routinely measured laboratory & radiological parameters: coagulation parameters and liver function as key prognosticators

TL;DR

This study uses automated machine learning to predict recovery outcomes in severe aneurysmal subarachnoid hemorrhage patients using routine lab and imaging data.

Contribution

The study introduces a novel application of automated machine learning for outcome prediction in poor-grade aSAH patients using routinely measured clinical parameters.

Findings

Logistic regression models achieved perfect prediction for both 1-month and 3-month functional outcomes.

Key predictors included coagulation parameters, liver function markers, and targeted temperature management.

Decision tree algorithms provided actionable cut-points for clinical decision-making.

Abstract

The objective of this study was to develop and evaluate automated machine learning (aML) models for predicting short-term (1-month) and medium-term (3-month) functional outcomes [Modified Rankin Scale (mRS)] in patients suffering from poor-grade aneurysmal subarachnoid hemorrhage (aSAH), using readily available and routinely measured laboratory and radiological parameters at admission. Data from a pilot non-randomized trial of 60 poor-grade aSAH patients (Hunt-Hess grades IV or V) were analyzed. Patients were evenly divided between targeted temperature management (TTM) and standard treatment groups. The current state-of-the-art for aML was adopted to employ nine ML algorithms with hyperparameter tuning to develop algorithmic models predicting 1 month and 3-months mRS scores. Model performance was evaluated using macro-weighted average Area Under the Receiver Operating Curve (mWA-AUROC) analysis and additional metrics. Logistic regression algorithmic models achieved perfect prediction (mWA-AUROC = 1, accuracy = 100%, sensitivity and specificity = 100% [95% CI: 83.16 − 100%]) for both 1-month and 3-month mRS outcomes. For 1-month outcomes, neutrophil count, platelet count, and gamma-glutamyl transferase levels were identified as key predictors. For 3-month outcomes, patient gender, activated partial thromboplastin time, and serum aspartate aminotransferase levels were most impactful. Decision tree algorithms (mWA-AUROC = 0.9-0.925) identified specific cut-points for various parameters, providing actionable information for clinical decision-making. Positive prognostic factors included alkaline phosphatase levels higher than mid-value of their normal range, absence of hydrocephalus, use of targeted temperature management (TTM), and specific cut-offs for coagulation and liver function parameters. The use of TTM was reinforced as a key prognosticator of mRS outcomes at both time points. We have made our developed models and the associated architecture available at GitHub. This study demonstrated the potential of aML in predicting functional outcomes for poor-grade aSAH patients. The identification of novel predictors, including liver function and coagulation parameters, opens new avenues for research and intervention. While the perfect predictive performance warrants cautious interpretation and further validation, these models represent a step towards personalized medicine in aSAH management, potentially improving prognostication and treatment strategies.

Figures