# Early mortality risk prediction in severe fever with thrombocytopenia syndrome using an interpretable machine learning model based on routine clinical parameters

## TL;DR

A machine learning model using six routine clinical parameters helps predict early mortality in severe fever with thrombocytopenia syndrome, offering a practical tool for timely care.

## Contribution

A high-performance, interpretable machine learning model for SFTS prognosis using only six routine clinical parameters.

## Key findings

- The model achieved an AUC of 0.960 in training and 0.938 in internal validation.
- Age and lactate dehydrogenase were the strongest predictors of mortality.
- The model maintained strong performance in two external validation cohorts.

## Abstract

Severe Fever with Thrombocytopenia Syndrome (SFTS) is characterized by high mortality and rapid progression, necessitating accurate early prognosis to optimize supportive care. However, current predictive tools often lack interpretability, require sophisticated tests unavailable in resource-limited areas, or suffer from poor generalizability. This study aimed to develop an interpretable, parsimonious, and deployable machine learning model for early mortality prediction in SFTS.

We analyzed data from 834 SFTS patients across three medical centers in Anhui, China. A LightGBM model was developed using a derivation cohort (n = 571) and validated on internal (n = 143) and two independent external cohorts (n = 80 and n = 183). Model interpretability was enhanced using SHapley Additive exPlanations (SHAP), and a web-based calculator was deployed for clinical use.

The LightGBM model identified six routine clinical parameters—Age, Lactate Dehydrogenase (LDH), Activated Partial Thromboplastin Time (APTT), Uric Acid (UA), Creatinine (CRE), and Body Temperature—as the most influential predictors. Integrating these features, the model achieved robust discrimination with an Area Under the Curve (AUC) of 0.960 in the training set and 0.938 in the internal validation set. Crucially, it maintained strong performance in two independent external validation cohorts (AUC 0.871 and 0.877). SHAP analysis revealed that Age and LDH were the strongest risk factors, while Temperature exhibited a non-linear relationship with mortality risk.

We developed and validated a high-performance, interpretable ML model for SFTS prognosis relying on only six readily available parameters. By deploying this parsimonious model as an online calculator, we provide a practical decision-support tool to facilitate early risk stratification and timely intervention, particularly in resource-limited settings.