A Novel Machine Learning Approach for Predicting Prognosis of SFTS Patients in the Early Stages of Disease.

Objective: Severe fever with thrombocytopenia syndrome (SFTS) is an emerging tick-borne disease characterized by high morbidity and mortality rates. Timely detection and prognosis prediction are critical for implementing effective clinical interventions. This study aimed to develop a binary classification machine learning model utilizing early clinical and laboratory indicators to predict the prognosis of SFTS patients, facilitating early clinical decision-making. Methods: We conducted a retrospective study including 233 SFTS patients diagnosed from October 2021 to May 2024. Clinical and laboratory data at initial diagnosis were collected and subjected to baseline analysis and correlation analysis to identify significant indicators. Using the area under the receiver operating characteristic curve as an indicator of model performance, select the analytical model among machine learning (LR) models, XGBOOST, LightGBM, and random forest. A binary classification machine learning model for predicting survival outcomes was constructed using a logistic regression algorithm in conjunction with identified metrics. The model's performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). The model was externally validated using a separate cohort of 73 patients. Result: A total of 233 patients were included in this study, among whom 146 (62.7%) survived and 87 (37.3%) died, with 208 assigned to the internal cohort (177 to the training set and 31 to the test set) and the external validation cohort consisted of 73 patients, including 52 survivors (71.2%) and 21 mortality cases (28.8%). Based on the AUC, the LR model (0.750) is selected, where the values of XGBOOST, LightGBM, and random forest are 0.649, 0.661, and 0.716, respectively. The logistic regression model, incorporating age, lactate dehydrogenase (LDH), albumin, activated partial thromboplastin time (APTT), and platelet count, demonstrated robust predictive accuracy with an AUC of 0.913 (95% CI: 0.821-1.000) for the test set and 0.923 (95% CI: 0.842-1.000) for the external validation set. The model exhibited an accuracy of 0.863, with a sensitivity of 0.952 and an NPV of 0.977. The specificity and PPV were 0.827 and 0.690, respectively. These results indicate that the model maintains robust discriminative performance in an independent cohort, suggesting its potential utility as a screening tool with high sensitivity and favorable NPV. Conclusion: The model constructed using the first diagnostic indicators described above can accurately determine the patient's prognosis, which can help the clinician intervene earlier as well as take the necessary life-saving measures, and has the potential to improve the survival rate of SFTS patients.