Interpretable machine learning for the prediction of death risk in patients with acute diquat poisoning

Machine Learning


This study effectively predicted the risk of death in patients with acute DQ poisoning using interpretable machine learning methods and common clinical indicators. As the use of PQ has gradually decreased, the incidence of poisoning by its substitute herbicide, DQ, has gradually increased. Currently, there is no specific antidote for DQ poisoning; therefore, the death rate of poisoned patients is highFourClinicians face great challenges in both the assessment and clinical treatment of such poisoning. Therefore, obtaining a simple and intuitive assessment method is highly important for quickly identifying the risk of death in acute and critical patients with rapid DQ intoxication.

DQ is a potent redox cycler that is readily converted to a free radical, which, when reacted with molecular oxygen, generates superoxide anions and, subsequently, other redox products. These products can induce lipid peroxidation in cell membranes and potentially lead to cell death18When DQ enters the body, it is reduced by receiving a single electron from NADPH, which is the primary source of reducing equivalents in cells, forming NADP+ and a highly unstable DQ. In turn, DQ transfers an electron to molecular oxygen (O2) to generate O2+.

DQ can revert to its initial state and undergo this continuous process to generate large quantities of O2.This O2 is subsequently neutralized spontaneously or through superoxide dismutase (SOD) activity, resulting in the formation of hydrogen peroxide (H2O2) and O219Under normal circumstances, H2O2 is converted to water through the action of catalase and glutathione peroxidase. However, in the presence of a substantial increase in reactive oxygen species production, the defense mechanisms within the cell, such as nonenzymatic constituents or antioxidant enzymes, are overburdened, leading to oxidative stress. Consequently, cellular dysfunction and injury occur.20,21,22.

DQ is believed to significantly affect hepatic and renal toxicity through the involvement of free radicalstwenty oneThis compound specifically induces damage to the kidney by affecting its excretory function, leading to conditions such as oliguria, anuria, proteinuria, haematuria, pyuria, azotemia, acute renal failure, and acute tubular necrosis.23,24In this study, consistent with previous findings, renal impairment was found to be a risk factor for death in patients with acute DQ poisoning. At the same time, DQ can also damage the liver, central nervous system, lungs, etc., as well as damage to the local reproductive system and the skin have also been reported.3,4,25Dyspnoea, pulmonary oedema, and respiratory depression are manifestations of pulmonary injury. However, unlike for PQ poisoning, there are no reports of pulmonary fibrosis caused by DQ poisoning.26,27In fact, in animal experiments, DQ caused mild and reversible damage to type I alveolar epithelial cells but not to type II alveolar epithelial cells.28Currently, there are no known remedies or successful treatments for DQ poisoning, and the focus of treatment has been on minimizing absorption and/or improving elimination18,29.

This study is the first to apply machine learning to predict the risk of death from acute DQ poisoning. Machine learning models are widely used in clinical diagnostics, precision treatments, and health monitoring and have achieved good results30,31Each model has its own advantages and disadvantages. For example, Random Forest has the benefit of fewer predictor variable assumptions than traditional modelling strategies and has minimal overfitting compared to simple classification and regression trees. However, Random Forest model has the fundamental issue of being a black box model. When alarms sound, medical staff are unsure of what immediate action to take until the patient is checked (cannot describe relationships within data32In this study, we employed machine learning combined with SHAP to assess the risk of death in patients with acute DQ poisoning. Previous studies were primarily relied on logistic regression analysis and have not yet explored the application of machine learning. Consequently, there remains a dearth of evidence regarding the benefits of machine learning in predicting the risk of death in patients with DQ poisoning. Our results demonstrate that all four models exhibit strong performance, with Random Forest surpassing traditional logistic regression analysis in terms of efficiency, as indicated by the ROC curves. We further plotted the importance features of random forest. The results revealed that Cr, PaCO2DQd, lactic acid, and WBC were important features for predicting death in patients with acute DQ poisoning. Higher levels of Cr, lactic acid, oral dosage of DQ, and WBC were associated with an increased risk of death, while lower levels of PaCO2 were also correlated with a greater risk of death. Most poisoning cases are related to the intentional ingestion of concentrated liquid formulations. In this study, the results showed a direct relationship between DQ intake and patient death. With the increase in the oral dose of DQ, the death rate of patients increased significantly, consistent with the results of previous studies33 that have shown that the ingestion of more than 15 mL of a rapid dose of 20% concentrated formulation of DQ is usually fatal.

The results of this study showed that the higher the lactic acid concentration, the greater the risk of death. The prognostic ability of arterial lactate levels has been assessed in various critical care patient groups, including those with septic shock, circulatory shock, recent surgical procedures, burns, and trauma. The level of lactic acid has emerged as a reliable predictor of mortality in individuals with severe illness.34In previous studies on the prognosis of acute PQ poisoning, clinical cases from different countries have shown that lactic acid is a good predictive factor.35,36.

In this study, a lower PaCO2 suggested a greater risk of death. In one study, a decrease in PaCO2 caused cerebral vasoconstriction, with a 1 mmHg change in PaCO2 Corresponding to a decrease in cerebral blood flow of 1.8 mL/100 g/min37According to the results, the WBC count is associated with poor outcomes. Many toxic diseases, such as acute organophosphate insecticide poisoning (AOPP), increase the WBC count, making it a poor indicator of prognosis.38In previous studies, in patients with acute PQ poisoning, an elevated WBC count was one of the indicators of poor prognosis.39.

The random forest model had a higher F1-score, accuracy, AUC, and MCC, and the Brier score was also the lowest. Compared to the other models, its overall performance was slightly better. DCA demonstrated that the four models provided a good net benefit within a range of thresholds (Fig. 1C). Overall, all the four models demonstrated good predictive performance, with Random Forest performing slightly better.

The SHAP calculation method was used in this study, which shows a list of important features, from most important to least important (from top to bottom). All the features contributed equally to the prediction of nonsurvival and survival, but the feature weights contained in the different models were not the same (Fig. 2). We provided two examples to illustrate the interpretability of the model, one for a nonsurviving patient and one for a surviving patient (Fig. 3). All four models presented very consistent predictive results in a straightforward manner, enabling clinicians to clearly observe the weights contributed by the included features in the model predictions. Individual predictors are greatly influenced by subjective factors; for example, the oral dose of patients is subjective and may not be very accurate, and vomiting dose, gastric lavage time, etc., affect the actual amount of absorption. Most earlier studies included only the patient's clinical test indicators and not their vital signs. This study combined objective indicators and patient status to objectively and intuitively evaluate the prognosis of patients with acute DQ poisoning.

Limitations

The sample size was small, which may have led to bias. In the future, we hope to continue to expand the sample size, summarize previous research experience, and strengthen the cooperation between basic and clinical studies to carry out high-quality clinical research for further demonstration. This research was based on a retrospective analysis; here, data were acquired from two distinct medical facilities, but due to limited data availability, the samples could not be divided into a testing group. Consequently, external validation is necessary to further evaluate the performance of our results.



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *