We collected data on patients with wild mushroom poisoning admitted to 5 county hospitals in Enshi, Hubei Province, China, between January 2010 and May 2022. Critical illness was defined as the occurrence of an admission to an intensive care unit, hemodialysis therapy, referral to a higher-level hospital, or death. We collected 24-hour data from 567 patients from 5 hospitals. We used data from 322 patients from 2 of the hospitals as the training set, of which 56 were critically ill and 245 were noncritically ill. Data from 245 patients from 3 other hospitals were used as the test set to assess the performance of our model, of which 43 were critically ill and 202 were noncritically ill.

We used the following inclusion criteria to develop the condition assessment model: (1) patients older than 14 years of age and (2) definite consumption of wild mushrooms before the onset of the disease. Data used in the diagnostic model were excluded when the following conditions were met: (1) age younger than 14 years; (2) possible concurrent consumption of other foods causing acute poisoning; and (3) previous cardiac, hepatic, and renal disorders, as well as mental disorders.

To ensure the reliability of the results and that model use characteristics were readily available, we excluded variables that omitted more than 30% of the observations. The missing values of continuous attributes were filled with the mean value of each attribute, and the missing values of discrete attributes were filled with the mode of each attribute. The values of each feature were normalized for the support vector machine and logistic regression.

Four popular machine learning classification algorithms, including extreme gradient boosting (XGBoost) [32], random forest, support vector machine, and logistic regression, were applied in this study to build the classification models. We implemented machine learning algorithms using Python (version 3.9; Python Software Foundation) and several Python modules (panda, numpy, scipy, sklearn, xgboost, shap, and matplotlib). Hyperparameter tuning was performed by a grid search based on 5-fold cross-validation to select the best area under the receiver operating characteristic curve (AUC) value for the classification models.

A proper model interpretation must be supplied for the machine learning model. Model predictions were interpreted using Shapley additive explanations (SHAP) [33,34]. SHAP is a model-independent interpretation technique that helps to interpret the results of prediction models. The interpretation is based on the SHAP value for each feature, which indicates the feature’s contribution to the risk of being predicted as critically ill. Having a positive SHAP value indicates that the corresponding feature contributes to a higher risk of the patient being critically ill and is a risk factor. On the other hand, having a negative SHAP value indicates that the corresponding feature contributes to a lower risk of the patient being critically ill and is a protective factor.

To validate the performance of the model, we compared the best machine learning model with the HOPE6 and TALK scoring models [7]. In addition, we included the results of each patient’s primary care physician’s assessment of the patient’s condition to explore the diagnostic performance of our model compared with that of the physician’s judgment.

We evaluated model performance in the test set by calculating (1) the AUC; (2) decision curve analysis (DCA); and (3) sensitivity, specificity, positive predictive value, negative predictive value, positive likelihood ratio, and negative likelihood ratio. The AUC is often used to assess the performance of various prediction models and is robust to category imbalance [35]. Based on the receiver operating characteristic curve, we chose the best predictive value (i.e., the value that is closest to the perfect model) for cases to fix the category imbalance for whether the condition is critical or not [36]. By calculating the AUC of different models, the discriminatory ability of different models can be compared. However, the AUC only focuses on the overall accuracy of the models and does not focus on the relationship between benefit and risk associated with different cutoff values in different models. DCA, on the other hand, permits the assessment of the range of threshold probabilities for a model to have value, the magnitude of the benefit, and the best model among numerous candidates. DCA figures out the “clinical net benefit” of one or more predictive models over a range of threshold probabilities. A threshold probability is a minimum chance that a disease needs further intervention, and the “clinical net benefit” takes into account the relative harms of false positives and false negatives [37].

This study was approved by the Ethical Committee of Renmin Hospital of Xianfeng, and informed consent was waived because this study was retrospective and used deidentified data (XFRY2021-12). The privacy and confidentiality of all individuals included in this study were strictly protected, and their data were used only for the purposes of this research.

The patient cohort participating in this study included data from 567 cases of mushroom poisoning of patients admitted to 5 county hospitals in the Enshi area. The case data included the following types of mushroom poisoning: gastroenteritis, neuropsychiatric symptoms, acute liver damage, acute renal failure, myocardial injury, and combined types [7]. For all data, the length of stay was 4.45 (95% CI, 3.45-5.44) and 3.52 (95% CI 3.21-3.82) days for critically ill and noncritically ill patients, respectively, and the cost of hospitalization was ¥11,113.09 (95% CI ¥8059.92-¥14,166.27; ¥1=US $0.14) and ¥2392.08 (95% CI ¥2188.34-¥2595.82), respectively.

Table 1 shows the baseline data of patients. The majority of the indicators did not differ significantly between the training set and the test set, and the remaining indicators (P<.01) were excluded from the machine learning model.

Among the algorithms, XGBoost achieved the highest AUC value, with AUCs of 0.83 (95% CI 0.77-0.90) and 0.90 (95% CI 0.83-0.96) in the internal and external validation sets (Figures 2 and 3). In the DCA (Figure 4), the XGBoost model had a greater net benefit compared to other methods over a wide range of threshold probabilities. As a result, we selected XGBoost as a suitable algorithm for developing the prediction model and conducted additional analyses to determine its predictive validity.

The feature ranking interpretation of the XGBoost model based on the SHAP algorithm (Figure 5) shows that lactate dehydrogenase (LDH), aspartate aminotransferase (AST), international normalized ratio (INR), serum sodium, alanine transaminase (ALT), hemoglobin, white blood cell, urea, total bilirubin (TBiL), creatine kinase-MB (CK-MB), creatinine, heart rate, indirect bilirubin (IBiL), and prothrombin time (PT) were important features of the XGBoost model. LDH and AST were the most influential factors, and their contribution was considerably greater than that of other indicators (Figure 5). Overall, the characteristics of LDH, AST, INR, serum sodium, ALT, hemoglobin, white blood cell, urea, CK-MB, creatinine, heart rate, and PT were positively correlated with the results and were risk factors; meanwhile, TBiL and IBiL were negatively correlated with the results and were protective factors (Figure 6).

In terms of the AUC, all the machine learning models performed significantly better than HOPE6, TALK, and physicians’ assessment (Figure 3). Moreover, across a wide range of threshold probabilities (or clinical preferences), all machine learning models performed better than HOPE6, TALK, and physicians’ assessment (Figure 4). In terms of AUC and DCA, XGBoost was the top model (Figures 3 and 4).

The sensitivity, specificity, positive predictive value, negative predictive value, positive likelihood ratio, and negative likelihood ratio of each model are shown in Table 2. Almost all patients were classified as critically ill in the HOPE6 and TALK scores (specificity: 0.00, 95% CI 0.00-0.02 for HOPE6 and 0.07, 95% CI 0.04-0.11 for TALK). In the test set, the XGBoost model had a sensitivity of 0.93 (95% CI 0.81-0.99) and a specificity of 0.79 (95% CI 0.73-0.85), whereas the physicians’ assessment had a sensitivity of 0.86 (95% CI 0.72-0.95) and a specificity of 0.66 (95% CI 0.59-0.73). The results indicate that the current diagnostic model can more precisely evaluate a patient’s condition, assist physicians in formulating more precise treatment plans, lessen the harm caused by incorrect treatment, and reduce patients’ treatment expenses.

Mushroom poisoning is a global food safety event. Early triage of patients with mushroom poisoning is essential for the formulation of treatment options and reduction of mortality. The objective of this study was to develop a machine learning–based triage model that could assess whether a patient with mushroom poisoning was critically ill within 24 hours of admission to support clinical decision-making. To our knowledge, this is the first time a machine learning algorithm has been used for the early triage of mushroom poisoning. The research demonstrates that the model developed using the XGBoost algorithm is superior to previous methods for triaging critically ill patients with mushroom poisoning. In addition, we discovered that liver dysfunction had the greatest impact on the model (more than 50%), with LDH and AST being the 2 most influential factors.

In this paper, we compared 4 machine learning models, 2 scoring models (HOPE6 and TALK), and clinical experts’ assessment results and found that machine learning models outperformed conventional methods. First, machine learning algorithms are entirely data driven, whereas scoring models and physicians’ evaluations are based on expert knowledge. Second, machine learning algorithms can learn and infer nonlinear higher-order connections between clinical factors and patient outcomes. Scoring models have the advantage of being simple to calculate and interpret; however, for complex clinical episodes (e.g., progression to critical illness), they may be overly simplistic in assuming that the severity of a patient’s condition is a linear combination of multiple factors.

XGBoost, a cutting-edge tree-based gradient boosting method, allowed us to create more accurate predictive models than other machine learning models; consequently, it was selected as our final model. The XGBoost model had greater sensitivity (0.93, 95% CI 0.81-0.99) and specificity (0.79, 95% CI 0.73-0.85) than physicians’ assessment (sensitivity: 0.86, 95% CI 0.72-0.95; and specificity: 0.66, 95% CI 0.59-0.73). As expert consensus, the HOPE6 and TALK models have not been validated by clinical data, and because the identification and treatment of critically ill patients have important clinical implications, scoring models based on expert experience will overestimate the condition of every patient. In this study, the HOPE6 and TALK models have extremely high sensitivity (1.00, 95% CI 0.92-1.00) and extremely low specificity (HOPE6: 0.00, 95% CI 0.00-0.02; and TALK: 0.07, 95% CI 0.04-0.11). Consequently, they are able to identify all critical cases and play an important role in clinical practice, allowing critical patients to be treated and the mortality rate to be as low as possible. However, an excessive number of noncritically ill patients will be misidentified as critically ill patients, which may result in the waste of medical resources and physical harm to patients as a result of overtreatment. Therefore, our model has greater advantages for the identification of critically ill patients and may aid physicians in making assessments. Moreover, the model’s ability to adjust cutoff values provides greater flexibility and more nuanced insights into the patient’s condition, making it a useful tool for physicians. The application software created by our current model is available on GitHub [38].

The study identified 14 clinical variables, of which AST, ALT, LDH, INR, PT, TBiL, IBiL, urea, creatinine, and CK-MB were consistent with current clinical evidence. These factors were classified as hepatic (AST, ALT [39-41], LDH [42], TBiL [39], and IBiL [43]), coagulation (INR [39,41] and PT [43]), renal (urea [43] and creatinine[40,44]), and cardiac impairment (CK-MB [42]). Using the SHAP method, we found that the indicators characterizing liver function impairment contributed 57.5% to the prediction results of the model and had the greatest impact. Interestingly, the study also identified several clinical indicators that lack relevant clinical evidence, such as serum sodium, hemoglobin, and heart rate, highlighting the potential for machine learning to identify novel relationships between clinical factors and patient outcomes.

Despite the promising results, the study has several limitations. The machine learning method is data driven, and the model’s performance is dependent on the quality and completeness of the data. Additionally, the study cohort was from a single location in China, limiting the model’s applicability to other regions. Future studies should collect data from more locations to enhance the model’s robustness and generalizability.

Using routinely accessible data at the time of triage, we found that machine learning has good predictive performance in the triage of patients with mushroom poisoning. Compared with other machine learning algorithms, clinical guidelines, and clinician assessment, the XGBoost model has better diagnostic performance and can be used to select core indicators for a triage model. Machine learning may be a powerful prognostic indicator for early warning in critically ill patients, which has a significant impact on the triage of patients, formulation of treatment options, and the allocation of medical resources.