JMIR Publications

ABSTRACT

Background:

Mushroom poisoning is a global food safety event. Early triage of patients with mushroom poisoning is essential for selecting the best treatment plan and reducing mortality. The current clinical studies about mushroom poisoning were either nongeneric based on a few clinical samples or empirically based on the literature and expert knowledge. Machine learning models can provide an objective way based on data for the triage of patients.

Objective:

In this study, we aimed to develop a machine learning model to rapidly and accurately identify critically ill patients.

Methods:

A total of 567 patients were collected from 5 primary care hospitals and facilities in Enshi, Hubei, and divided into two groups. A total of 322 patients from 2 hospitals were used as the training cohort, and 245 patients from 3 hospitals were used as the test cohort. The machine learning model was developed based on the eXtreme Gradient Boosting (XGBoost) algorithm, and its performance was compared with that of conventional critical condition assessment methods. Performance was assessed using the area under the receiver operating characteristic curve (AUC), decision curve, sensitivity, and specificity. Feature contributions were evaluated using SHapley Additive exPlanations (SHAP).

Results:

A model developed using XGBoost to assess the condition of patients with wild mushroom poisoning within 24 hours contained 11 clinical indicators. The model was compared with the HOPE6 (AUC=0·50 (95% confidence interval (CI) 0·40-0·60), sensitivity=1·00 (1·00-1·00), specificity=0·00 (0·00-0·00)) and TALK (AUC=0·53 (0·44-0·63), sensitivity=1·00 (1·00-1·00), specificity=0.07 (0·03-0·10)) disease assessment models; clinical expert assessment (AUC=0·76 (0·67-0·85), sensitivity=0·86 (0·76-0·96), specificity=0.66 (0·60-0·73)) was also compared to the XGBoost model, which had better performance (AUC=0·88 (0·81-0·95), sensitivity=0·93 (0·85-1·00), specificity=0·67 (0·61-0·74)). The decision curve analysis (DCA) showed that the XGBoost model had a greater net benefit compared to other methods over a wide range of threshold probabilities.

Conclusions:

We successfully established the model for the early disease assessment of mushroom poisoning. The model can rapidly and accurately identify critically ill patients and will possibly serve as an important basis for the selection of treatment options and referral of patients, potentially reducing patient mortality and improving clinical outcomes.