A machine learning approach for mortality prediction only using non-invasive parameters.

Area of Science:

• * Computational biology and medical informatics.
• * Development of predictive models for critical care outcomes.

Background:

• * Traditional mortality prediction relies on laboratory measurements, posing challenges in resource-limited settings.
• * Lack of access to professional laboratorians and equipment hinders early mortality prediction in rural areas.
• * Need for efficient, accurate, and applicable mortality prediction methods in remote locations.

Purpose of the Study:

• * To develop a novel machine learning-based mortality prediction method using only non-invasive parameters.
• * To improve the efficiency, accuracy, and applicability of early mortality prediction, especially in remote areas.
• * To compare the performance of machine learning models against traditional scoring methods.

Main Methods:

• * Development of a new feature selection method based on Bayes error rate.
• * Training of four machine learning models (including LightGBM) using non-invasive parameters.
• * Comparison of machine learning models with and without laboratory data against traditional scoring systems.

Main Results:

• * LightGBM achieved the highest accuracy (0.797) and AUC (0.879) using only non-invasive parameters.
• * Machine learning models showed no significant performance difference with or without laboratory measurements.
• * Reduced feature sets (≤50) still allowed machine learning models to outperform traditional systems (AUC > 0.83).

Conclusions:

• * Machine learning models utilizing non-invasive parameters offer excellent mortality prediction performance.
• * These models are comparable to methods requiring laboratory measurements.
• * The proposed approach is highly applicable for mortality risk evaluation in rural areas and remote battlefields.