Prediction of Pralidoxime Dose in Patients with Organophosphate Poisoning Using Machine Learning Techniques

authors:

avatar Babak Mostafazadeh 1 , avatar Shahin Shadnia 1 , avatar Sayed Masoud Hosseini 1 , avatar Mitra Rahimi ORCID 1 , avatar Haleh Talaie 2 , avatar Seyed Ali Mohtarami 3 , * , avatar Peyman Erfan Talab Evini 1 , avatar Joshua King 4

Department of Clinical Toxicology, Toxicological Research Center, Excellence Center of Clinical Toxicology, Loghman Hakim Hospital, Shahid Beheshti University of Medical Sciences, Tehran, Iran
Toxicological Research Center, Loghman-Hakim Hospital, Shahid Beheshti University of Medical Sciences, Tehran, Iran
Department of Forensic Medicine, Legal Medicine Research Center, Legal Medicine Organization, Tehran, Iran
Maryland Poison Center, University of Maryland Medical Center, Baltimore, USA
Health Scope: Vol.In Press, issue In Press; e143897
published online: April 24, 2024
article type: Research Article
received: December 20, 2023
revised: January 29, 2024
accepted: February 24, 2024
DOI: https://doi.org/10.5812/healthscope-143897

how to cite: Mostafazadeh B, Shadnia S, Hosseini S M, Rahimi M, Talaie H, et al. Prediction of Pralidoxime Dose in Patients with Organophosphate Poisoning Using Machine Learning Techniques. Health Scope. 2024;In Press(In Press):e143897. https://doi.org/10.5812/healthscope-143897.

Abstract

Background:

Early and appropriate antidotal therapy is crucial for patients with organophosphate poisoning.

Objectives:

Given the lack of a comprehensive consensus on the optimal dose of pralidoxime for patients with organophosphate poisoning, this study aims to develop a machine learning-based prediction model to determine the individualized pralidoxime dose for these patients.

Methods:

The dataset was divided into training and test sets with a 70:30 ratio. Feature selection was conducted using Pearson’s correlation coefficient (filter approach) method. Both classification and regression were employed to develop the prediction model using the selected features. The performance of the developed models was evaluated using ten-fold cross-validation and various metrics, including sensitivity, specificity, accuracy, F1-score, and AUC. The models were implemented and assessed using the scikit-learn library in Python.

Results:

After applying exclusion criteria, data from 325 patients were utilized to train and test the machine-learning models. In the classification approach, the random forest method achieved superior performance with an AUC of 98.6. In the regression approach, the gradient boosting regressor, with an R2 value of 65.4, outperformed other algorithms. Feature selection revealed that muscular weakness, plasma cholinesterase activity, and blood urea nitrogen were the most significant predictors of pralidoxime dose in the classification model. In the regression model, the top predictors were age, HCO3-VBG, and atropine bolus. Many of the selected features coincide with those identified in previous studies, with muscular weakness being particularly significant in both models.

Conclusions:

The most effective algorithms could be employed to develop a clinical decision support system for personalized pralidoxime dosage prediction in patients with organophosphorus poisoning. However, the study is constrained by its small sample size, retrospective design, and the absence of an external validation cohort. Conducting a prospective multicenter study with a larger sample size is crucial to validate the findings of this study.

1. Background

Organophosphate (OP) poisoning is a major global health issue caused by exposure to OP compounds, which are widely used as pesticides. Organophosphates are among the most common causes of accidental, occupational, and suicidal intoxications due to their widespread use, easy accessibility, and simplicity in synthesis (1). Despite the role of pesticides in enhancing agricultural productivity, concerns are increasing about their adverse effects on human health (2). Organophosphate poisoning is more prevalent in developing countries where highly hazardous pesticides (HHPs) are more readily available. Significant incidences of intentional or unintentional OP poisoning occur in the agricultural regions of developing countries in South Asia and Southeast Asia, including China, India, Sri Lanka, Iran, and the Maldives (2, 3). Annually, the global death toll from OP poisoning is estimated to range from 250 000 to 350 000 (4). In rural areas, poisoning with organophosphorus compounds typically occurs deliberately with suicidal intent (5). One study reported that 51.7% of OP poisonings were deliberate, while 21.7% were accidental (4). Previous studies have indicated that intentional self-poisoning is associated with significantly higher mortality than unintentional poisoning (6). Given the high prevalence of intentional poisoning with organophosphorus compounds, particularly in developing countries, and the high mortality associated with it, implementing necessary therapeutic interventions for these patients is crucial.

Organophosphates inhibit the enzyme acetylcholinesterase (AChE) by phosphorylating the serine hydroxyl group on the enzyme. Acetylcholinesterase is responsible for hydrolyzing the neurotransmitter acetylcholine (ACh) at cholinergic synapses. Therefore, AChE inhibition leads to the accumulation of ACh at the cholinergic synapse, resulting in prolonged stimulation of nicotinic and muscarinic ACh receptors, which manifests as OP toxidrome (1, 7). In addition to supportive care, the standard treatment includes three main therapeutic approaches: Atropine to block muscarinic receptors, oxime-type reactivators (e.g., pralidoxime, obidoxime) to dephosphorylate the inhibited AChE and benzodiazepines for seizure management.

In Iran, pralidoxime chloride (2-PAM) is the oxime of choice and is commonly used as an antidote for OP poisoning. It works by reactivating AChE, thus reducing the accumulation of ACh (8). For adults, the initial recommended dose of pralidoxime chloride typically ranges from 1 to 2 grams, followed by a continuous infusion at a rate of 250 to 500 mg per hour. However, dosing guidelines can vary based on the specific product used and the severity of the poisoning. Close monitoring of the patient's clinical response is crucial to determine the need for additional doses or adjustments in the infusion rate. In pediatric patients, the dose of pralidoxime is generally calculated based on body weight, with an initial dose typically ranging from 20 to 50 mg/kg, followed by a continuous infusion at a rate of 5 to 10 mg/kg per hour (9-11).

There is a divergence of opinion regarding the effectiveness of pralidoxime in treating OP poisoning; some studies have reported that its administration leads to a decrease in mortality among patients poisoned with organophosphorus agents (12-15), while others indicate no effect (16, 17). Several previous studies examining the effectiveness of various dosages of pralidoxime are outlined in Table 1 (11, 14, 16, 18-25). The appropriate dose of pralidoxime for a specific individual depends on several factors, including the severity of poisoning, age, weight, and individual response (26). Additionally, some studies recommend high-dose pralidoxime (12), while others have reported that high doses may be associated with a higher mortality rate and recommend low-dose pralidoxime (27). With the expansion of information technology in many research fields, including healthcare, real-world studies such as non-interventional and observational studies have become significant data sources for clinical research (28). Compared to traditional models, machine learning approaches offer immense benefits in handling real-world evidence. Unlike traditional models, which cannot handle complex, interacting, high-dimensional variables, machine learning models are more accurate and provide greater generalizations (29). The use of machine learning techniques based on real-world research has become popular recently, with examples including models for predicting tacrolimus blood concentration in patients with autoimmune diseases, the prediction of vancomycin dose through XGBoost, and the prediction of warfarin maintenance dose through LightGBM (30-32).

Table 1.

Characteristics of Previous Studies Investigating the Effectiveness of Pralidoxime in Patients with Organophosphates Poisoning a

AuthorsStudyCountry, YearNumber of Patients SexAge (y)Pralidoxime DosageOutcome
Pawar et al. (14)RCTIndia, 2006200, poisoning by anticholinesterase pesticideMale; control group: 52; study group: 57Control group: 29 (22 - 35); Study group: 28 (22 - 33)Control group: 2 g loading dose over 30 min, then a bolus dose of 1 g/4 h for 48 h; study group: 2 g loading dose over 30 min, then a constant infusion of 1 g/h for 48 h.Patients in the study group required less atropine, a short duration of ventilatory support, and less frequency of intubation.
Syed et al. (16)RCTIndia, 2015100 patients with OP poisoningSex ratio (male:female); control group: 1:1.56; study group: 1:1.62Control group: 28.2 ± 9.9; study group: 29.1 ± 10.9Control group: Received I.V saline; study group: (30 mg/kg loading dose over 30 min followed by 8 mg/kg/h continuous infusion for a maximum of 7 daysThere was no statistically significant difference between the two groups in terms of mortality, hemodynamic parameters, atropine requirements, duration of ventilation, and ICU stay
Eddleston et al. (11)RCTSri Lanka, 2009235 patients with OP poisoningMale; control group: 92 (80.7); study group: 96 (79.3)Control group: 29.5 (23 - 42); study group: 31 (22 - 48)Control group: Normal saline infusion; study group: 2 g loading dose over 20 min, then a constant infusion of 0.5 g/h until 7 daysMortality was non-significantly higher in patients receiving pralidoxime, and the need for intubation was similar in both groups
Thunga and Pandey (18)Cross-sectional, nonrandomized observational studyIndia, 2013256 OPs poisoned patientsSex ratio (male: female): 2.3:1Majority of patients were in the age group of 21 - 30 yearsControl group: Normal saline infusion; study groups: Intermittent (1 g/q8h), continuous infusion (500 mg/h), continuous infusion (1 g/h)The incidence of intermediate syndrome, number of ventilation days, total atropine requirement, number of hospitalization days, and mortality rate significantly reduced in continuous infusion of pralidoxime at 500 mg/hour
Due (19)Comparative studyVietnam, 2014Control group: 54; study group: 108Male; control group: 30 (55.6); study group: 64 (59.3)Control group: 25.5 ± 10; study group: 29.5 ± 14.2Control group: Received a mean total dose of 7.2 ± 4.1 g; study group: Received a mean total dose of 20.0 ± 12.7 gIn the study group, patients received significantly lower doses of atropine and required a shorter duration of hospital stay, and the mortality rate was lower than the control group
Mahesh et al. (20)Randomized open-labeled prospective studyIndia, 2013Control group: 45; study group: 37Male:female, control group: 35:10; study group: 21:16Control group: 30.1 ± 7.3; study group: 31.3 ± 8.9Control group: 2 g bolus followed by 1g/6 h; study group: 2 g bolus followed by 8 mg/kg/h infusionDuration of mechanical ventilation, mean dosage of atropine administered, and incidence of intermediate syndrome were significantly lower in the study group than in the control group
Chaudhary et al. (21)Case-control studyIndia, 2013Control group: 35; study group: 35Male:female, control group: 25:10; study group: 24:11Control group: 24.80 ± 8.3; study group: 25.17 ± 9.2Control group: -; study group: 30 mg/kg body weight in 200 mL of normal saline over 30 min followed by 2.0 g in 200 mL of normal saline over 30 min, at 6 h intervals for initial 72 hThere was no statistically significant disparity observed between the two groups in relation to the total amount of atropine administered, necessity for intubation, average length of hospitalization, and mortality rate.
Cherian et al. (22)RCTIndia, 2005Control group: 11; study group: 10NANAControl group: Normal saline infusion; Study group: PAM infusion of 12 g/day for 3 days in severe cases and 4 g/day for 3 days in moderate casesThere was no difference between the treatment and placebo groups in relation to the necessity for intubation, average length of hospitalization, and mortality rate.
Cherian et al. (23)RCTIndia, 1997Control group: 55; Study group: 55Male, control group: 34 (62); study group: 41 (75)Control group: 26.5 ± 10.3; study group: 28 ± 10.1Control group: Normal saline infusion for 3 days; study group: Infusion of 12 g over 3 daysThe requirement of ventilator support (37 vs. 22), the incidence of an intermediate syndrome (36 vs. 19), and the mortality rate (16 vs. 3) were higher in the study group than in the control group
Banerjee et al. (24)Open-label, parallel-group clinical studyIndia, 2011Control group: 30; study group: 30Male, control group: 11 (37); study group: 14 (47)Control group: 34.3 ± 8.8; study group: 34.6 ± 9.8Control group: -; study group: a dose of 0.5 - 1 g/6 hThe mortality rate, requirement of ventilator, and duration of hospital stay in the two groups failed to show any statistically significant difference.
Banerjee et al. (25)Open-label, parallel-group, randomized clinical trialIndia, 2014Control group: 60; study group: 60Male, control group: 26 (43); study group: 23 (38)Control group: 34.3 ± 8.8; study group: 34.6 ± 9.8Control group: -; study group: A dose of 1 g every 6 hours for a period of 5 daysPralidoxime therapy did not offer any appreciable benefit over atropine alone in terms of reducing mortality and ventilator requirement. Patients in the study group experienced longer duration of hospital stay

a Values are expressed as No. (%), mean ± SD or median (IQR) unless otherwise indicated.

2. Objectives

In this study, we aimed to predict an optimal pralidoxime therapeutic dose based on real-world evidence, including initial symptoms, clinical presentation, vital signs, and laboratory parameters at presentation, using machine learning methods to enhance its efficacy in patients poisoned by OPs.

3. Methods

3.1. Study Population

The roadmap of the proposed system for predicting pralidoxime dose in patients with OP poisoning is illustrated in Figure 1. This retrospective study was conducted at Loghman Hakim Hospital, Tehran, Iran's primary referral center for poisoned patients, from March 2016 to April 2021. The study included patients aged 14 years and older who had been poisoned with OPs and met specific criteria. These criteria included documented OP exposure and a plasma cholinesterase (PChE) activity of less than 4300 U/L upon admission, indicative of acute OP poisoning (33). A total of 325 patients were included, particularly those who received commonly recommended pralidoxime dose regimens. These regimens typically involved an initial intravenous infusion of 1 - 2 g (10 - 20 mg/mL) over 15 - 30 minutes, followed by a continuous infusion of 500 mg/h until 24 hours after the patient was weaned from atropine. The study aimed to predict the outcome of dried pulmonary secretions and adequate oxygenation following pralidoxime administration. Exclusion criteria included non-receipt of pralidoxime, multidrug toxicity, and the presence of severe chronic comorbidities. Ethical approval for the study was obtained from the ethics committee (IR.SBMU.RETECH.REC.1401.257), and informed consent was obtained from all participants.

Graphical abstract: The roadmap of the proposed system to predict pralidoxime dose in patients with organophosphate poisoning
Graphical abstract: The roadmap of the proposed system to predict pralidoxime dose in patients with organophosphate poisoning
Figure 1.

Graphical abstract: The roadmap of the proposed system to predict pralidoxime dose in patients with organophosphate poisoning

3.2. Collection and Processing of Data

Data for this study were obtained from electronic medical records. Two researchers reviewed the medical records of the patients, and a questionnaire was used to extract clinical information from the electronic databases of Loghman Hakim Hospital. The questionnaire covered various patient demographics, medical history, symptoms, and laboratory test results, as well as hospital-related factors such as the need for intubation and duration of hospitalization. The dataset underwent several preparation processes, including the removal of rows with missing values greater than 70%, application of min-max scalar and standard scalar, data validation, under-sampling, and splitting of the dataset.

3.3. Feature Selection

Feature selection in machine learning involves identifying the most relevant attributes or parameters to improve model performance. In this research, the filter approach using Pearson’s correlation coefficient was employed to select features by assessing the correlation between each variable and the predictive outcome. The aim was to identify the key predictors of the total pralidoxime dose necessary for maintaining sufficient oxygenation. Only variables demonstrating a strong correlation with the predictive outcome, such as dried pulmonary secretions and adequate oxygenation, were retained through the feature selection process. Variables with zero correlation coefficients were excluded from the feature subset, while those with high coefficients were included in the selected variables subset.

3.4 Statistical Analysis

Based on the outcomes of the Kolmogorov-Smirnov and Shapiro-Wilk tests, it was determined that all continuous variables exhibited non-normal distributions. Consequently, these variables were represented by their median values and interquartile ranges and were analyzed using the Kruskal-Wallis and Mann-Whitney U tests. Categorical variables were depicted in terms of absolute frequencies (n) and were analyzed using the chi-square test.

3.5. Model Development and Performance Evaluation

The prediction model for pralidoxime dose in patients poisoned by OPs was assessed using two methods: (1) regression, with pralidoxime dose considered as a continuous variable; and (2) classification, with pralidoxime dose divided into two classes, low dose, and high dose. In the low dose class, patients received total doses of less than 14 000 g, and in the high dose class, patients received total doses of 14 000 g or more. For the first method, 25 machine learning models were used, and for the second method, 15 machine learning models were trained. Models with the best performance are listed in Table 2 (the top ten models for the regression method) and Table 3 (the top 14 models for the classification method). A ten-fold cross-validation technique was employed to train and test the machine learning algorithms using both complete and selected feature datasets. This method involves dividing the dataset into ten sections and performing the holdout method ten times. The final dataset was randomly split into training (70%) and testing (30%) sets. Model creation, hyper-parameter tuning, and model performance evaluation were conducted using these datasets. This approach prevents random data bias and ensures an equal distribution of data between the training and test sets. It is important to note that the testing dataset was not used during the training process. Hyperparameters were tuned using the training dataset through the cross-validation method. Python software was used to develop a model utilizing classification models. The performance of the classification algorithms was assessed by testing them on the testing dataset after training. Five commonly used efficiency testing metrics, including sensitivity, specificity, accuracy, F1-score, and AUC, were utilized to evaluate the effectiveness of classification algorithms in predicting the pralidoxime dose in OP-poisoned patients. These performance measures were then used to compare the performance of each trained classifier with that of other machine-learning systems. The efficiency evaluation metrics for the classifiers are as follows:

Accuracy=TP+TNTP+YN+FP+FN×100
Recall aka sensitivity=TPTP+FN×100
Presicion=TPTP+FP×100
f-measure=2Presicion×RecallPresicion+Recall

Four performance measures were employed in the regression method to evaluate the performance of each algorithm: Mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and R-squared (R2). The calculations for these metrics are as follows:

MAE=1n∑j=1nyj- y^j
MSE=1n∑j=1n(yj- y^j)2
RMSE=1n∑j=1n(yj-y^j)2

In the process of regression analysis, we employed the standardization technique to scale the variables. Standardization involves centering the values around the mean and adjusting them to have a unit standard deviation. This is achieved by dividing the variable by its standard deviation after subtracting the mean, as indicated by the following equation:

XStandard=X-mean(X)Standard Deviation(X)
Table 2.

Results of Ten-Fold Cross-Validation for Regressors Performance on Selected Predictors After Feature Scaling in Regression Method

IndexModelMAEMSERMSER2RMSLEMAPE
GbrGradient boosting regressor9208.1856265695067.524515210.52680.65400.72100.8991
light gumLight gradient boosting machine11082.3208411516903.582918637.13180.58180.86531.2540
EtExtra trees regressor9964.7046390583726.461317809.58350.48840.72870.9194
AdaAdaBoost regressor14937.2844428589688.691619027.89170.48371.11432.2290
xgboostExtreme gradient boosting9210.9117342642130.792117211.78120.41990.70870.9305
RfRandom forest regressor11670.8163433264351.126219619.44160.41450.78121.0514
HuberHuber regressor16389.36481018552916.341127720.56790.17421.15061.2774
ParPassive aggressive regressor16079.87721110867086.391129187.17240.09231.13191.0646
EnElastic net19619.83891026946561.562229350.46100.00351.20722.2364
DtDecision tree regressor10381.8142832112670.286624221.97960.00170.82610.7595
Table 3.

Results of Ten-Fold Cross-Validation for Classifiers Performance on Selected Predictors in Classification Method

Index ModelAccuracyAUCRecallPrecisionF1Kappa
rfRandom forest classifier0.91520.98650.90610.93080.91690.8307
light gumLight gradient boosting machine0.90730.97560.91590.91290.91070.8145
gbcGradient boosting classifier0.90710.95840.93180.8960.91250.8136
adaAda boost classifier0.88970.92180.8970.89680.89450.7786
etExtra trees classifier0.87980.96190.8720.90130.8830.7599
xgboostExtreme gradient boosting0.87630.95480.88260.8810.8790.7528
ldaLinear discriminant analysis0.82750.85150.83790.83140.83280.6544
ridgeRidge classifier0.79660.00.82120.79420.80440.5924
dtDecision tree classifier0.78320.78390.78790.79670.78290.5672
lrLogistic regression0.7480.82540.78710.7510.76510.4947
qdaQuadratic discriminant analysis0.70690.80360.64770.8260.66350.4168
knnK neighbors classifier0.69470.79090.70150.72250.70430.3898
nbNaive bayes0.68560.78430.76820.68560.71950.363
SvmSVM - linear kernel0.57960.00.57050.67660.53750.1577

4. Results

4.1. Patients

After a comprehensive review of electronic medical records, a total of 5,380 patients diagnosed with OP poisoning were identified within the database, forming the primary cohort. Among them, 983 patients had incomplete medical records, and 4,072 patients were excluded based on predetermined exclusion criteria. Specifically, 383 patients had a normal range of PChE, 1,057 patients had multidrug toxicity, 937 patients had chronic comorbidity, 529 patients were admitted more than 24 hours after poisoning, 549 patients were below 14 years of age, and 617 patients did not receive pralidoxime. Ultimately, 325 patients were enrolled in the study. Figure 2 depicts the patient selection process. The descriptive characteristics of the study sample are presented in Table 4.

Patient selection flowchart
Patient selection flowchart
Figure 2.

Patient selection flowchart

Table 4.

The Study Sample's Descriptive Characteristics a

VariablesPralidoxime DoseTotal (n = 325)P-Value
Low Dose (n = 172)High Dose (n = 153)
Sex 0.06
Male9669165
Female7684160
Age33 (25)31 (22)-0.092
Time to hospitalization 4 (10)4 (7)-0.332
Sialorrhea0.000
No160135295
Yes121830
Bradycardia0.370
No169152321
Yes314
Rales0.000
No153111264
Yes171153324
Bronchospasm0.467
No11697213
Yes5656112
Intubation0.763
No121105226
Yes514899
Incontinence0.138
No166142308
Yes61117
Fasciculation0.000
No160122282
Yes123143
Muscular weakness0.000
No14147188
Yes31106137
Paralysis0.000
No172140312
Yes01313
Tachycardia0.738
No8875163
Yes8478162
Confusion0.270
No9896194
Yes7457131
Lethargy0.180
No170153323
Yes202
Coma0.380
No127106233
Yes454792
Agitation0.017
No172148320
Yes055
Seizures0.621
No167147314
Yes5611
ICU admission0.284
No9575170
Yes7778155
The amount of toxin consumed 90 (120)150 (150)1000.00
GCS 15 (7)15 (7)12.50.589
Cholinesterase level first 305 (971)317 (1114)316.50.787
Cholinesterase level last458 (835)715 (1898)583.500.172
PH-VBG 7.35 (0.11)7.34 (0.13)7.3450.291
PCO2-VBG38.4 (14.4)41.5 (11.7)40.30.001
HCO3-VBG 19.6 (4.7)20.4 (5.2)20.40.346
Cr 1.2 (0.3)1.2 (0.3)1.20.005
BUN28 (18.5)33 (15)310.002
AST26 (16.3)25 (12)250.595
ALT 19 (13)17 (9)180.017
ALP204 (79)178 (79)1920.002
Bulos of atropine0.5 (1)1 (1)0.50.593

a Values are expressed as No. or median (IQR).

4.2. Feature Selection

Significant and relevant features for predicting the pralidoxime dose in OP-poisoned patients were selected using Pearson’s correlation coefficient (filter approach). Figure 3 shows the top ten variables selected for the classification method, where muscular weakness, PChE activity, and blood urea nitrogen (BUN) scored the highest for predicting pralidoxime dose. In the regression method, after scaling the variables using the standardization technique, the top ten features in predicting the dose of pralidoxime from most to least important were age, HCO3-VBG, bolus of atropine, aspartate aminotransferase (AST), pulmonary rale, PH-VBG, PChE activity, muscular weakness, BUN, and alanine transaminase (ALT) (Figure 4).

Variables are selected by feature selection in the classification method.
Variables are selected by feature selection in the classification method.
Figure 3.

Variables are selected by feature selection in the classification method.

Variables are selected by feature selection using the regression method.
Variables are selected by feature selection using the regression method.
Figure 4.

Variables are selected by feature selection using the regression method.

4.3. Performance of Prediction Models

Performance metrics were calculated to compare the effectiveness of prediction models in classification and regression methods. Table 2 presents the results of the top ten prediction algorithms for the regression method using selected standard features to predict the pralidoxime dose in patients with OP poisoning. The gradient-boosting regressor emerged as the most accurate and responsive model, exhibiting the highest R2 value. This prediction model had an MAE of 9208.19, an MSE of 265695067.52, an RMSE of 15210.52, and an R2 of 65.4. Figure 5 depicts the learning curves of the regression method before (A) and after (B) feature scaling, demonstrating improved convergence of training and cross-validation scores after scaling. Table 3 summarizes the results of the top 14 prediction algorithms for the classification method using selected features to predict the pralidoxime dose in OP-poisoned patients. Based on ten-fold cross-validation findings, the random forest classifier showed the highest accuracy and sensitivity, as well as the highest AUC value among the classifiers tested. This model achieved a mean accuracy of 91.5%, a recall or sensitivity of 90.6%, a precision of 93.1%, an F1-score of 91.7%, Kappa statistics of 83.1%, and an AUC of 98.6. Figure 6 presents the confusion matrix (A) and learning curve (B) for this model, indicating a good fit as the training and cross-validation curves converge at a high score.

The learning curve of the regression method before (A) and after (B) feature scaling
The learning curve of the regression method before (A) and after (B) feature scaling
Figure 5.

The learning curve of the regression method before (A) and after (B) feature scaling

The confusion matrix (A) and learning curve (B) for the classification method
The confusion matrix (A) and learning curve (B) for the classification method
Figure 6.

The confusion matrix (A) and learning curve (B) for the classification method

5. Discussion

Immediate antidotal treatment in patients with severe OP poisoning is crucial. The first line of antidotal therapy for OP poisoning includes atropine and pralidoxime (34). In antidote therapy, prescribing the correct dose of the antidote is vital, although the appropriate dose of pralidoxime in patients with OP poisoning remains controversial, with various dosage schedules being employed (20, 35). Several studies have investigated the ideal dose of pralidoxime in OP-poisoned patients, but the results have been inconclusive. For instance, one study reported that a pralidoxime dose of 20.0 ± 12.7 g was more effective than a dose of 7.2 ± 4.1 g, leading to a reduction in the duration of hospitalization and mortality rate (19). Another study found that a pralidoxime dose of 2 g followed by an 8 mg/kg/h infusion was more effective than a dose of 2 g followed by 1 g every 6 hours (20). We hypothesize that an individualized pralidoxime dose, determined based on clinical presentations and laboratory parameters, might be helpful in these cases.

Machine learning techniques enable healthcare providers and researchers to leverage large amounts of patient data to make precise and individualized dose predictions. By tailoring treatment plans to the unique characteristics of each patient, machine learning can contribute to improved treatment outcomes, reduced adverse effects, and enhanced patient safety. Machine learning algorithms can analyze complex patterns and relationships within patient data to make accurate dose predictions (36). For instance, in one study, researchers used machine learning methods to develop a predictive model for determining the appropriate dosage of lapatinib in patients with metastatic HER2(+) breast cancer, utilizing real-world data and applying both machine learning and deep learning techniques (29). Another study reported an optimal dosing algorithm for vancomycin in patients with MRSA and other gram-positive bacterial infections, developed using machine-learning methods (31). Asiimwe et al. assessed the performance of 21 machine-learning algorithms in predicting stable warfarin doses in sub-Saharan Black-African patients (37).

In this study, we applied a machine learning technique for the first time to predict pralidoxime doses in patients poisoned by OPs based on demographic profiles, clinical symptoms, and laboratory parameters. We employed two different methods for this purpose. The first method, classification, grouped patients into two categories based on the pralidoxime dose: Either low dose (< 14000 g) or high dose (≥ 14000 g). In the second method, after feature scaling using the standardization technique, the pralidoxime dose was included in the machine-learning model as a continuous variable. A total of 35 features were selected for both the classification and regression methods. The following features were among the top ten in both methods: Age, muscular weakness, ALT, AST, BUN, and initial PChE level. Apart from these shared features, in the classification method, the last measured level of PChE (last measurement within the first 24 hours of hospitalization), ALP, PCO2-VBG, and the amount of consumed toxin was among the top ten predictors of pralidoxime. In the regression method, additional top predictors included HCO3-VBG, received a bolus of atropine, pulmonary rale, and PH-VBG.

Many of the features selected in this study overlap with those identified in our previous study, which examined predictors of OP poisoning severity (38). The AChE enzyme prevents the buildup of the neurotransmitter ACh at various muscarinic and nicotinic sites in the body by hydrolyzing it. This enzyme has both a serine site and an anionic site. The OP molecule attacks the serine site located within the active site of the enzyme. As a result, through an irreversible interaction, the serine site is phosphorylated, and a strong covalent bond is formed, leading to the inactivation of the enzyme's active site. Pralidoxime functions as a reactivator of AChE by attaching to the enzyme's anionic site, close to the previously attached OP molecule at the serine site. The pralidoxime molecule sacrifices itself by becoming phosphorylated instead of the enzyme since it has a higher affinity for phosphorylation by OP agents than the enzyme's serine site. Consequently, the OP molecule detaches from the enzyme and forms an OP-pralidoxime complex, which is then hydrolyzed. This process results in enzyme reactivation (39). The main effect of pralidoxime is the restoration of ACh esterase at nicotinic sites in the body, alleviating symptoms such as muscle weakness, fasciculations, and paralysis (40). According to our results, muscular weakness was one of the most significant features in predicting the dose of pralidoxime, selected in both the classification and regression methods. Other predictors of pralidoxime dosage are also features directly related to the severity of organophosphorus poisoning and can be used to predict the dosage of pralidoxime in these patients.

In the classification method among different machine learning algorithms, the random forest classifier demonstrated the strongest performance, with an AUC value of 98.6%. In the regression method, the gradient-boosting regressor showed the strongest performance, with an R2 value of 0.65. Both the random forest classifier and the gradient boosting regressor are decision-tree-based machine learning algorithms. Algorithms utilizing decision trees demonstrate high efficacy in predictive modeling, especially for small-to-medium structured datasets with substantial missing data (38).

The novelty of this study lies in the utilization of machine learning methodology to forecast the optimal dosage of pralidoxime for individuals affected by OP poisoning. This investigation marks the inaugural development of a machine learning-based predictive model for pralidoxime dosage, employing two distinct approaches: Classification and regression. The strength of the study lies in the use of machine learning methods, which offer numerous advantages such as enhanced accuracy, efficiency, and decision-making capabilities. Nevertheless, the study is subject to several limitations, notably the small sample size and the retrospective nature of the data, which led to a considerable amount of missing data in the dataset. Additionally, the absence of an external validation cohort and the recruitment of patients solely from a single hospital further constrain the study. Thus, to obtain confirmatory results, a more robust multicenter study with a larger sample size is imperative.

5.1. Conclusions

Machine learning-based prediction algorithms can be utilized to anticipate the appropriate dosage of pralidoxime for patients suffering from organophosphorus (OP) poisoning. Through the application of feature selection techniques, age, HCO3-VBG, received bolus of atropine, and AST were identified as the most significant predictors of pralidoxime dosage in the regression method. The gradient-boosting regressor emerged as the most effective algorithm with the highest predictive performance. In the classification method, muscular weakness, the last measured level of PChE, BUN, ALP, and AST were determined to be the most crucial predictors of pralidoxime dosage, with the random forest classifier demonstrating the best predictive performance. These algorithms have the potential to contribute to the development of a clinical decision support system for the precise individualized dosing of pralidoxime in OP poisoning cases. Nevertheless, further research with a larger sample size is imperative to substantiate these findings.

Acknowledgements

References

  • 1.

    Lorke DE, Petroianu GA. The Experimental Oxime K027-A Promising Protector From Organophosphate Pesticide Poisoning. A Review Comparing K027, K048, Pralidoxime, and Obidoxime. Front Neurosci. 2019;13:427. [PubMed ID: 31191210]. [PubMed Central ID: PMC6547910]. https://doi.org/10.3389/fnins.2019.00427.

  • 2.

    Eizadi-Mood N, Mahvari R, Akafzadeh Savari M, Mohammadbeigi E, Feizi A, Mirmoghtadaei P, et al. Acute pesticide poisoning in the central part of Iran: A 4-year cross-sectional study. SAGE Open Med. 2023;11. [PubMed ID: 36778200]. [PubMed Central ID: PMC9909076]. https://doi.org/10.1177/20503121221147352.

  • 3.

    Bonvoisin T, Utyasheva L, Knipe D, Gunnell D, Eddleston M. Suicide by pesticide poisoning in India: a review of pesticide regulations and their impact on suicide trends. BMC Public Health. 2020;20(1):251. [PubMed ID: 32075613]. [PubMed Central ID: PMC7031890]. https://doi.org/10.1186/s12889-020-8339-z.

  • 4.

    Razwiedani LL, Rautenbach P. Epidemiology of Organophosphate Poisoning in the Tshwane District of South Africa. Environ Health Insights. 2017;11:1178630217694150. [PubMed ID: 28469445]. [PubMed Central ID: PMC5345965]. https://doi.org/10.1177/1178630217694149.

  • 5.

    Samimi A, Rahmani AH, Ababaf R, Zeidooni L. An Investigation of Clinical Symptoms and Treatment of Organophosphate Poisoning among Patients Referred to Razi Hospital during 2006 - 2012. Asia Pac J Med Toxicol. 2016;5(4):107-10. https://doi.org/10.22038/apjmt.2016.8135.

  • 6.

    Eddleston M. Patterns and problems of deliberate self-poisoning in the developing world. QJM. 2000;93(11):715-31. [PubMed ID: 11077028]. https://doi.org/10.1093/qjmed/93.11.715.

  • 7.

    Balali-Mood M, Balali-Mood K. Neurotoxic disorders of organophosphorus compounds and their managements. Arch Iran Med. 2008;11(1):65-89. [PubMed ID: 18154426].

  • 8.

    Gupta R, Parmar M. Pralidoxime. StatPearls [Internet]. Treasure Island, FL: StatPearls Publishing; 2023.

  • 9.

    Eddleston M, Buckley NA, Eyer P, Dawson AH. Management of acute organophosphorus pesticide poisoning. Lancet. 2008;371(9612):597-607. [PubMed ID: 17706760]. [PubMed Central ID: PMC2493390]. https://doi.org/10.1016/S0140-6736(07)61202-1.

  • 10.

    Balali-Mood M, Balali-Mood K, Hosseini Shirazi F. Recent Advances In Treatment of Acute Organophosphorous Nerve Agents Poisoning. Iran J Pharm Res. 2006;2:79-87.

  • 11.

    Eddleston M, Eyer P, Worek F, Juszczak E, Alder N, Mohamed F, et al. Pralidoxime in acute organophosphorus insecticide poisoning--a randomised controlled trial. PLoS Med. 2009;6(6). e1000104. [PubMed ID: 19564902]. [PubMed Central ID: PMC2696321]. https://doi.org/10.1371/journal.pmed.1000104.

  • 12.

    Eyer P, Buckley N. Pralidoxime for organophosphate poisoning. Lancet. 2006;368(9553):2110-1. [PubMed ID: 17174692]. [PubMed Central ID: PMC1978066]. https://doi.org/10.1016/S0140-6736(06)69843-7.

  • 13.

    Balali-Mood M, Shariat M. Treatment of organophosphate poisoning. Experience of nerve agents and acute pesticide poisoning on the effects of oximes. J Physiol Paris. 1998;92(5-6):375-8. [PubMed ID: 9789840]. https://doi.org/10.1016/S0928-4257(99)80008-4.

  • 14.

    Pawar KS, Bhoite RR, Pillay CP, Chavan SC, Malshikare DS, Garad SG. Continuous pralidoxime infusion versus repeated bolus injection to treat organophosphorus pesticide poisoning: a randomised controlled trial. Lancet. 2006;368(9553):2136-41. [PubMed ID: 17174705]. https://doi.org/10.1016/S0140-6736(06)69862-0.

  • 15.

    Thiermann H, Worek F. Pro: Oximes should be used routinely in organophosphate poisoning. Br J Clin Pharmacol. 2022;88(12):5064-9. [PubMed ID: 35023196]. https://doi.org/10.1111/bcp.15215.

  • 16.

    Syed S, Gurcoo SA, Farooqui AK, Nisa W, Sofi K, Wani TM. Is the World Health Organization-recommended dose of pralidoxime effective in the treatment of organophosphorus poisoning? A randomized, double-blinded and placebo-controlled trial. Saudi J Anaesth. 2015;9(1):49-54. [PubMed ID: 25558199]. [PubMed Central ID: PMC4279350]. https://doi.org/10.4103/1658-354X.146306.

  • 17.

    Kharel H, Pokhrel NB, Ghimire R, Kharel Z. The Efficacy of Pralidoxime in the Treatment of Organophosphate Poisoning in Humans: A Systematic Review and Meta-analysis of Randomized Trials. Cureus. 2020;12(3). e7174. https://doi.org/10.7759/cureus.7174.

  • 18.

    Thunga G, Pandey S. Effectiveness of Different Dosage Regimen of Pralidoxime in Organophosphorus Poisoning in Tertiary Care Hospital. Value Health. 2013;16(7). A678. https://doi.org/10.1016/j.jval.2013.08.1988.

  • 19.

    Due P. Effectiveness of High Dose Pralidoxime for Treatment of Organophosphate Poisoning. Asia Pac J Med Toxicol. 2014;3(3):97-103. https://doi.org/10.22038/apjmt.2014.3377.

  • 20.

    Mahesh M, Gowdar M, Venkatesh CR. A Study on Two Dose Regimens of Pralidoxime in the Management of Organophosphate Poisoning. Asia Pac J Med Toxicol. 2013;2(4):121-5. https://doi.org/10.22038/apjmt.2013.2033.

  • 21.

    Chaudhary SC, Singh K, Sawlani KK, Jain N, Vaish AK, Atam V, et al. Prognostic significance of estimation of pseudocholinesterase activity and role of pralidoxime therapy in organophosphorous poisoning. Toxicol Int. 2013;20(3):214-7. [PubMed ID: 24403730]. [PubMed Central ID: PMC3877488]. https://doi.org/10.4103/0971-6580.121669.

  • 22.

    Cherian MA, Roshini C, Visalakshi J, Jeyaseelan L, Cherian AM. Biochemical and clinical profile after organophosphorus poisoning--a placebo-controlled trial using pralidoxime. J Assoc Physicians India. 2005;53:427-31. [PubMed ID: 16124349].

  • 23.

    Cherian AM, Peter JV, Samuel J, Jaydevan R, Peter S, Joel S, et al. Effectiveness of P2AM (PAM-pralidoxime) in the treatment of organophosphorus poisoning (OPP) a randomized, double blind placebo controlled clinical trial. J Assoc Physicians India. 1997;45(1):22-4.

  • 24.

    Banerjee I, Tripathi SK, Sinha Roy A. A Study on Comparative Evaluation of Add-on Pralidoxime Therapy over Atropine in the Management of Organophosphorus Poisoning in a Tertiary Care Hospital. JK Sci. 2011;13(2):65-9.

  • 25.

    Banerjee I, Tripathi SK, Roy AS. Efficacy of pralidoxime in organophosphorus poisoning: revisiting the controversy in Indian setting. J Postgrad Med. 2014;60(1):27-30. [PubMed ID: 24625936]. https://doi.org/10.4103/0022-3859.128803.

  • 26.

    Hilmas E, Hilmas CJ. Medical Management of Chemical Toxicity in Pediatrics. In: Gupta RC, editor. Handbook of Toxicology of Chemical Warfare Agents. London: Academic Press; 2015. p. 1003-34. https://doi.org/10.1016/b978-0-12-800159-2.00068-3.

  • 27.

    Peter JV, Cherian AM. Organic insecticides. Anaesth Intensive Care. 2000;28(1):11-21. [PubMed ID: 10701030]. https://doi.org/10.1177/0310057X0002800102.

  • 28.

    Blonde L, Khunti K, Harris SB, Meizinger C, Skolnik NS. Interpretation and Impact of Real-World Clinical Data for the Practicing Clinician. Adv Ther. 2018;35(11):1763-74. [PubMed ID: 30357570]. [PubMed Central ID: PMC6223979]. https://doi.org/10.1007/s12325-018-0805-y.

  • 29.

    Yu Z, Ye X, Liu H, Li H, Hao X, Zhang J, et al. Predicting Lapatinib Dose Regimen Using Machine Learning and Deep Learning Techniques Based on a Real-World Study. Front Oncol. 2022;12:893966. [PubMed ID: 35719963]. [PubMed Central ID: PMC9203846]. https://doi.org/10.3389/fonc.2022.893966.

  • 30.

    Zheng P, Yu Z, Li L, Liu S, Lou Y, Hao X, et al. Predicting Blood Concentration of Tacrolimus in Patients With Autoimmune Diseases Using Machine Learning Techniques Based on Real-World Evidence. Front Pharmacol. 2021;12:727245. [PubMed ID: 34630104]. [PubMed Central ID: PMC8497784]. https://doi.org/10.3389/fphar.2021.727245.

  • 31.

    Huang X, Yu Z, Wei X, Shi J, Wang Y, Wang Z, et al. Prediction of vancomycin dose on high-dimensional data using machine learning techniques. Expert Rev Clin Pharmacol. 2021;14(6):761-71. [PubMed ID: 33835879]. https://doi.org/10.1080/17512433.2021.1911642.

  • 32.

    Liu Y, Chen J, You Y, Xu A, Li P, Wang Y, et al. An ensemble learning based framework to estimate warfarin maintenance dose with cross-over variables exploration on incomplete data set. Comput Biol Med. 2021;131:104242. [PubMed ID: 33578070]. https://doi.org/10.1016/j.compbiomed.2021.104242.

  • 33.

    Dong N, Wang S, Li X, Li W, Gao N, Pang L, et al. Prognostic nomogram for the severity of acute organophosphate insecticide self-poisoning: a retrospective observational cohort study. BMJ Open. 2021;11(5). e042765. [PubMed ID: 34031108]. [PubMed Central ID: PMC8149305]. https://doi.org/10.1136/bmjopen-2020-042765.

  • 34.

    Yung SZ, Cheong K, Lee JH. Preventive General Pediatrics. In: Yazdani S, Kamzan A, Kulkarni D, Lerner C, Newcomer CA, editors. General Pediatrics Board Review, Medical Specialty Board Review. New York: Oxford University Press; 2020. p. 1-43. https://doi.org/10.1093/med/9780190848712.003.0001.

  • 35.

    Peter JV, Moran JL, Graham P. Oxime therapy and outcomes in human organophosphate poisoning: an evaluation using meta-analytic techniques. Crit Care Med. 2006;34(2):502-10. [PubMed ID: 16424734]. https://doi.org/10.1097/01.ccm.0000198325.46538.ad.

  • 36.

    Habehh H, Gohel S. Machine Learning in Healthcare. Curr Genomics. 2021;22(4):291-300. [PubMed ID: 35273459]. [PubMed Central ID: PMC8822225]. https://doi.org/10.2174/1389202922666210705124359.

  • 37.

    Asiimwe IG, Blockman M, Cohen K, Cupido C, Hutchinson C, Jacobson B, et al. Stable warfarin dose prediction in sub-Saharan African patients: A machine-learning approach and external validation of a clinical dose-initiation algorithm. CPT Pharmacometrics Syst Pharmacol. 2022;11(1):20-9. [PubMed ID: 34889080]. [PubMed Central ID: PMC8752108]. https://doi.org/10.1002/psp4.12740.

  • 38.

    Hosseini SM, Rahimi M, Afrash MR, Ziaeefar P, Yousefzadeh P, Pashapour S, et al. Prediction of acute organophosphate poisoning severity using machine learning techniques. Toxicology. 2023;486:153431. [PubMed ID: 36682461]. https://doi.org/10.1016/j.tox.2023.153431.

  • 39.

    Gupta R, Parmar M. Pralidoxime. StatPearls [Internet]. Treasure Island, FL: StatPearls Publishing; 2022.

  • 40.

    Tush GM, Anstead MI. Pralidoxime continuous infusion in the treatment of organophosphate poisoning. Ann Pharmacother. 1997;31(4):441-4. [PubMed ID: 9101007]. https://doi.org/10.1177/106002809703100411.