Using a Machine Learning Approach to Predict Snakebite Envenoming Outcomes Among Patients Attending the Snakebite Treatment and Research Hospital Kaltungo, Northeastern Nigeria

1. Introduction

Snakebite envenoming, a neglected tropical disease, has emerged as a serious public health crisis in Nigeria and sub-Saharan Africa. Despite its devastating impact, including high morbidity and mortality rates among vulnerable populations and thousands of cases reported annually, it remains largely overlooked [1,2]. Globally, the situation reflects similar neglect, with the World Health Organization (WHO) estimating 81,000 – 138,000 deaths annually due to snake envenoming - a figure that is likely underestimated due to significant underreporting [3].

Historically, epidemiological investigation into snakebites, particularly those caused by Echis spp. (saw-scaled or carpet viper) in the Nigerian savanna region, revealed that bite incidence peaked during the rainy season, which coincided with peak farming activities [4,5]. Comprehensive data collected from hospitals in Wukari, Bambur, Zaria, Kaltungo, and Gombe, highlighted both long-term and seasonal patterns of these incidents and their impacts on local populations [4,5]. It was not until the 1980s that the true burden of snakebite envenoming became evident with reports documenting high morbidity and mortality rates in Northern Nigeria [6,7]. The 1990s saw a surge in research efforts with notable studies [5,8] highlighting the importance of snakebite envenoming as a significant public health problem. Building on these findings, Warrel et al. [9] emphasized the urgent need for improved data collection, robust research methodologies, and evidence-based policy decisions in combatting this issue.

Recently, there has been a significant shift towards adopting more rigorous and comprehensive approaches to studying snakebite epidemiology in Nigeria. These approaches include advanced methodologies such as systematic reviews, meta-analyses and mathematical modeling studies. Snakebite epidemiology in Nigeria is marked by a high incidence and mortality rate, with rural communities being disproportionately affected. [10,11]. Contributing factors include the country's geographic location, climate, settlements patterns and socioeconomic conditions [12].

The Snakebite Treatment and Research Hospital (SBTRH) in Kaltungo, originally known as the Snakebite Ward, has been a key center for the treatment and research on snakebite envenoming in sub-Saharan Africa for over two decades [5,13]. Handling an average of 2,500 cases annually, it is recognized as the largest snakebite hospital in sub-Saharan Africa, in terms of patient volume. Established in the 1990s, the hospital has played a pivotal role in reducing the burden of snakebite envenoming in the region, achieving high recovery rates [14]. The hospital's innovative approach, which integrates clinical care, research and community engagement, aligns closely with the WHO’s recommendation for comprehensive snakebite management [1].

Despite the high recovery rates observed at SBTRH, there is a significant lack of literature addressing the critical factors that predict patient recovery in this region. Identifying these factors is essential for optimizing medical practices and interventions related to snakebite envenoming management. Furthermore, the use of machine learning methods for predictive analytics in snakebite research remains underexplored. To bridge this gap, we piloted machine learning approaches to assess their potential in predicting snakebite envenoming outcomes. SBTRH provided an ideal setting for such analyses, offering a vast and high-quality dataset. We aimed to use this data to uncover patient characteristics contributing to the institution’s limited number of patients experiencing complications or poor outcomes.

Accordingly, our objective was to investigate patient characteristics as potential predictors of envenoming outcomes - specifically recovery, amputation, debridement, or death, among snakebite patients treated at SBTRH. To achieve this, we had three distinct aims: (1) to identify key differences in characteristics between patients experiencing recovery versus those experiencing amputation, debridement, or death; (2) to examine the strength of associations between patient characteristics and envenoming outcomes through logistic regression; and (3) to build an XGBoost classification model to predict envenoming outcomes based on patient characteristics. This comprehensive approach was designed to generate findings that could significantly influence clinical practice and decision making in the treatment of snakebite envenomation in SBTRH.

2. Materials and Methods

2.1. Data Collection

SBTRH utilizes a paper-based records system, with patient information documented in registers and folders provided by the Gombe State Hospital Services Management Board. For this single-center retrospective study, data from patient records between 1st January 2024 and 31st July 2024 were entered into a Microsoft Excel document. This time-period covered the dry season and onset of the rainy season. Ethical approval was obtained for this study as part of a larger data science capacity building project, granted by the Gombe State Ministry of Health.

Patient records provided the following information: month of snakebite occurrence, patient age, sex, state of origin, occupation, anatomical site of snakebite, species of snake, number of antivenom vials used, time (in hours) between snakebite and hospital presentation, and patient outcome (death, amputation, debridement, or recovery). The digitized data were cleaned and analyzed using R software (version 4.3.1). One variable, snake species, had missing data, as not every patient was able to identify the snake responsible for their bite.

2.2. Data Analysis

First, sociodemographic and clinical characteristics were summarized for all patients, as well as by treatment outcome. For continuous variables, a Kruskal-Wallis test was used, for categorical variables with cell counts greater or equal to five, a Chi-Squared test was used, and for categorical variables with cell counts less than five, the Fisher’s Exact test was used to evaluate differences in patient characteristics between treatment outcomes. Differences were considered to be statistically significant if the corresponding test’s p-value was less than 0.05, as per statistical convention.

Second, both univariate and multivariable logistic regression was used to explore associations between patient characteristics and patient outcomes. We grouped patients who experienced amputation, debridement, or death into one category, and compared them to patients who recovered. Further, we included broad age groups in this model (adult versus paediatric) instead of ten-year age groups, and we converted “number of hours between snakebite and presentation to hospital” into a categorical variable. For the latter, we used the median number of hours between snakebite and presentation to hospital to categorize this variable into two groups: less than the median, and greater than or equal the median. Finally, we considered a sub-analysis where we removed the variable for antivenom dose from both univariate and multivariable logistic regression models. This sub-analysis was undertaken because the causal pathway was not clear; a patient presenting with more severe symptoms, which impacts likelihood of recovery, may be given a high dose of antivenom. This may result in unexpected findings, such as antivenom appearing to promote poor outcomes rather than protect a patient from poor outcomes. We present the results of this sub-analysis in the main text, with additional results presented in the Supplementary Materials, where the antivenom variable was included in both regression models.

Third, we used a machine learning approach to evaluate how well patient characteristics were able to predict patient outcome. The same modified dataset that was used for the logistic regression analyses above was used for this as well. Accordingly, we have also provided additional results in the Supplementary Materials concerning the inclusion of the antivenom variable in machine learning models. The XGBoost method was chosen as a binary classification algorithm due to its ease of use, ability to deal with unbalanced classes, and ability to handle clinical data [15,16,17]. At its core, the XGBoost algorithm minimizes the error between observed and predicted data (through the loss function), and favors reduced model complexity, vis-a-vis regularization [18].

We randomly divided patients in our dataset; 70% of the patient records were used for training the XGBoost model, and 30% were used to test its performance. Classes with less than ten observations were removed from the dataset to ensure reasonable distribution of classes between the test and training dataset. We fixed the hyperparameter “scale_pos_weight”, as our outcome classes were not equal in size; far more patients recovered than those who experienced amputation, debridement, or death. As per convention, we set scale_pos_weight to be equal to the number of patients experiencing amputation, debridement or death, divided by the number of patients who recovered. We also fixed the learning rate at 0.01 to remain conservative. Few other hyperparameters were tuned with Bayesian Optimization, with cross validation over five folds [19]. Table 1 describes hyperparameters that were tuned, and corresponding ranges of values that were explored through Bayesian Optimization. These ranges were informed by hyperparameter values in similar studies [15,16,17,18].

Optimized hyperparameters were then used to build an XGBoost model. All features (patient characteristics) were used in the model as there were only eight features in total, however, a subset of features was subsequently selected based on statistical significance during the previous univariate logistic regression step. This simplified model was constructed using the same Bayesian Optimization approach described above. Both models’ performances were evaluated based on the area under the receiver-operator curve (AUROC) statistic. A high AUROC indicated better model performance as this optimized sensitivity and specificity in predicting patient outcome. To further evaluate the XGBoost models’ performances, their evaluation metrics were compared to those of a class-weighted random-forest algorithm and a logistic regression algorithm. Hyperparameters for both random-forest (mtry and ntrees) and logistic regression (mixture and penalty) were also tuned using a Bayesian Optimization approach. Of note, for random-forest, cross validation was not used as this model does not need it as a guard for overfitting, however, we sampled without replacement to account for unbalanced classes.

3. Results

As shown in Table 2, a total of 1,022 snakebite cases were seen in the hospital between 1^st January 2024 and 30^th June 2024. Among these, nearly two-thirds were 18 years of age or older (n = 638, 62%), and a majority were male (n = 734, 72%). Most snake bites occurred in April (n = 282, 28%), as this was the beginning of the rainy season. Roughly half of the patients were from Gombe State, which is where SBTRH is located (n = 498, 49%), while the other half of the patients came from neighbouring states, or Cameroon (grouped in with the “Other” category). Nearly half of the patients were farmers (n = 469, 46%). On average, patients presented to hospital within four hours of being bitten (interquartile range 2 – 8 hours). Most patients were bitten on either hand (n = 795, 78%), and most patients were bitten by carpet vipers (n = 810, 79%). Finally, most patients received a single dose of antivenom (n = 664, 65%), and most patients recovered from their snakebite (n = 931, 91%).

3.1. Differences in Patient Characteristics by Treatment Outcome

As illustrated in Table 3, there were statistically significant differences in patient characteristics by treatment outcome. Proportionally, there were more male patients among those who experienced amputation, debridement, or death (11%) compared to female patients (4.5%). Further, a proportionally higher number of adults experienced poor outcomes compared to paediatric patients (11% vs. 6%). Those who experienced amputation, debridement, or death had a longer time taken to arrive to hospital (median 5 hours) compared to those who recovered (median 4 hours). Finally, a proportionally higher number of patients who received two or more doses of antivenom experienced amputation, debridement, or death (18%), compared to those who received a single dose (7.2%) or no antivenom (2.9%).

3.2. Logistic Regression Analysis

Logistic regression analysis revealed key associations between patients’ characteristics and clinical outcomes (Table 4). According to univariate logistic regression, patients who arrived at SBTRH less than four hours after being bitten, female patients, as well as paediatric patients, were less likely to experience amputation, debridement, or death, compared to their counterparts. However, neither of these three patient characteristics remained statistically significantly associated with poor outcomes in the multivariable regression model. A higher antivenom dose was significantly associated with higher likelihood of experiencing poor outcomes in additional analyses (Table S1).

3.3. XGBoost Model

Finally, the patient data was divided into a training dataset (containing 691 patients), and a test dataset (containing 296 patients). The training dataset was used to build an XGBoost model. The scale_pos_weight hyperparameter was set at 9.85 to reflect the ratio of recoveries to amputations or debridement surgeries. Table 5 describes the variables that were used in both the full and simplified XGBoost models, as well as the hyper parameter values following Bayesian Optimization.

Similarly, a random-forest model was constructed with the full set of features (hyperparameter values mtry = 2 and ntrees = 20) and the simplified set of features (hyperparameter values mtry = 2 and ntrees = 11). A logistic regression model was also constructed with the full set of features (hyperparameter values mixture = 0 and penalty = 0) and the simplified set of features (hyperparameter values mixture = 0.53 and penalty = 0). The results of all models (on the test data) are shown in Table 6.

According to AUROC, which optimizes both sensitivity and specificity, all models had modest performance. AUROC values improved slightly when including the antivenom variable, as illustrated in Table S2. The XGBoost model had the highest AUROC value of all models when using the simplified set of features, however, random-forest and logistic regression models had comparable performance. While the AUROC of the simplified random-forest model was slightly lower in comparison to the other two models, it achieved perfect specificity, meaning a patient’s age, sex, and time taken to arrive to hospital could perfectly predict if they were going to experience amputation, debridement, or death. However, this came at a cost of having very low sensitivity, meaning these characteristics were unlikely to accurately predict recovery.

4. Discussion

This study reported on 1,022 cases of snakebite treated at SBTRH Kaltungo from January to June 2024, which we suggest is a representative number that provides insight into the epidemiology, treatment and outcomes of snakebite incidents seen in the facility that sees over 2,500 cases per year. The data revealed a critical pattern of patient demographics, clinical outcomes, and treatment modalities.

Our findings noted that a large proportion of the patients were adults (62%) and were predominantly male (72%), which is consistent with existing literature [20,21,22]. Further, the majority of bites occurred in April, which coincides with the onset of the rainy season in the region when farmers are clearing their farmlands and is consistent with global and local studies [23,24]. Gombe State, where SBTRH is located, accounts for 49% of cases which is consistent with recent studies about Gombe State contributing a higher percent of cases in the hospital than other states [25,26]. These other studies also noted a predominance of snakebites among male patients (72%) and adults (58%), with farmers (46%) being the most common occupation group. Such trends align with existing literature [27,28,29,30] that suggests that occupational hazards contribute to snakebite incidents, particularly in agricultural settings. In addition, the male to female ratio reflects societal norms where men are more likely to engage in outdoor labour thereby increasing their exposure to snakebites. This highlights the need for targeted public health interventions aimed at adult males, particularly in rural areas where agricultural activities predispose individuals to encounters with snakes. For example, interventions may be in the form of educational programmes to increase awareness about snakebite prevention during peak season.

Our study also noted that a majority of the bites occurred on the hand with 46% on the right hand and 36% on the left hand, which reflects occupational hazards where interactions with snakes occur during manual labour – particularly farming. The site of the bite is crucial in management as bites from the hands and feet are associated with severe local tissue damage due to its proximity to muscles, tendons and bones, which predisposes such patients to complications such as necrosis or secondary infection leading to debridement or amputation.

This study found that 79% of cases were attributed to carpet viper (Echis romani) which is consistent with previous studies [7]. It is a venomous snake known for its potent haemotoxic venom which causes severe coagulopathy and local tissue destruction [31]. Notably, 18% of our study’s patients were unable to identify the species of snake that bit them; this is common in snakebite reports, as relatives or patients may not have been able to identify the snake, or the snake may not have been killed and brought to the facility for identification. However, understanding the snake’s species is critical for appropriate and timely administration of antivenom, especially in areas where polyvalent antivenoms are not readily available. Developing a targeted community awareness campaign about risk to farmers and the importance of venomous snake identification would therefore be a beneficial intervention to curb the snakebite burden in this region, and will be a recommendation from our work.

We also noted that a significant majority of the patients (91%) recovered while 9% of cases either ended in amputation, debridement or death. Despite the high rate of recovery, we identified a trend that raised concern about the severity of some bites, particularly among males, adult patients, and patients who took more than four hours to arrive to hospital, all of whom were more likely to experience complications than their counterparts. Interestingly, these characteristics did not remain significantly associated with patient outcome in a multivariable model, but the association with hours between bite and hospitalization was close to statistical significance and may still be of clinical significance.

The higher proportion of male patients undergoing amputation or debridement or experiencing death may also reflect the occupational risks associated with farming, which is the most common occupation among male patients in this region [27,28]. This emphasizes the need for targeted and more aggressive interventions for these population groups, such as closer monitoring, early surgical intervention and careful repeated dosing of antivenoms. However, such actions cannot be implemented without early recognition of patients at risk for severe complications. Indeed, shorter time intervals between bite and hospital presentation as well as belonging to the paediatric age group were associated with better outcomes in univariate analysis. However, these trends did not remain significant in the multivariable model. This may suggest that other factors – including those that we did not extract from patient records – could have a stronger influence on patient outcomes. Nevertheless, the protective effect of early presentation aligns with existing literature [26,32], emphasizing the need for timely intervention in snakebite incidents.

The differences in rates of recovery between male and female patients, and adult and paediatric patients may also reflect in the physiological responses to envenomation, health seeking behaviour, or access to care. Additionally, a larger proportion of paediatric and male patients received a single dose of antivenom compared to adults and female patients. This raises an important question of whether clinical decision making regarding antivenom administration is influenced by age and sex, or whether these differences reflect underlying severity of envenomation. Further studies could investigate these patterns to underpin specific thresholds for antivenom dosing in relation to envenomation severity; this would ensure equitable and effective care across patient subgroups and different clinical stages of envenomation.

There is a dearth of literature that leverages machine learning methods for snakebite epidemiology, although these methods have been previously used for snake identification and geospatial modelling of snake habitats [33,34,35,36]. We sought to apply various machine learning approaches to our institutional dataset, as a proof-of-concept. Given the data we had, we were able to achieve AUROC values around 0.5, depending on the machine learning model used. The modest levels of model performance may be due to the relatively low proportion of deaths, amputations and debridement surgeries in our dataset. This led to a class imbalance in the data, affecting the sensitivity and specificity. Despite not having the highest AUROC, the random-forest model with a simplified set of features resulted in a perfect specificity value, but a very low sensitivity value. These results imply that age, sex, and time between bite and hospital presentation were sufficient to predict poor outcomes accurately. However, these features were insufficient to predict recovery. The highest AUROC value was achieved by the XGBoost model using a simplified set of features.

In general, for all models, performance tended to increase slightly when using a simplified set of features. This is of enormous clinical implication because the streamlined model can be implemented in a resource limited setting like SBTRH Kaltungo, where rapid, less resource-demanding and relevant decision-making is crucial to reduce snakebite morbidity and mortality. However, there was an enormous trade-off between model sensitivity and specificity. More studies of this nature are needed to build more robust machine learning models, as these can be a valuable tool in predicting patient outcomes. Identifying high risk patients early allows providers to adjust strategies, thereby improving recovery rates and reducing the need for invasive procedures such as debridement or amputation, or even mortality from snakebites. These findings may support the development of risk stratification tools that can guide clinical management of snakebite patients, thereby ensuring effective and efficient use of scarce resources and interventions like antivenoms.

The use of machine learning techniques to analyse large and clinically relevant datasets of snakebites and the application of Bayesian Optimization in tuning hyperparameters adds to the robustness of our findings and provides a valuable framework for future studies. Furthermore, the retrospective nature of the study design enabled us to explore a diverse array of patient characteristics and clinical outcomes, offering a comprehensive understanding of factors influencing snakebite prognosis.

However, our study was limited to single centre, which may affect the generalizability of the findings to other regions where snake species, patient demographics, healthcare infrastructure and type of antivenom may differ. To enhance the external validity of the findings of our studies, future studies should look at multicentre collaboration using a using a standardized protocol across different health care settings. Therefore, while our machine learning model demonstrated adequate performance in predicting clinical outcomes, further external validation is needed to improve the AUROC.

Future studies should aim at validating our findings and incorporating additional clinical and laboratory data in a multicentre study to enhance generalizability of the predictive models. Efforts should also be made at implementing machine learning-driven decision support tools in real time clinical settings to improve patient outcomes.

5. Conclusions

Conclusively, our study was able to identify several key predictors of clinical outcomes such as sex, age group, and time before presentation hospital. It also demonstrated that machine learning models such as XGBoost, random-forest, and logistic regression could be used to predict patient outcomes. These findings underscore the potential for the use of machine learning to enhance clinical decision making in resource limited settings.

The findings of this study have significant implications for public health practice. These include the development of targeted interventions for high-risk groups (i.e. males, farmers and individuals living in rural areas), optimization of antivenom administration for the high-risk patients (adult males who have a long travel-time to reach the hospital) and integration of machine learning-driven decision support tools into clinical settings. Furthermore, policymakers should prioritize resource allocation for snakebite prevention and treatment programmes, particularly targeting high-risk groups. Further validation of these findings is needed in a multi-centre study to support the integration of predictive models into routine clinical practice, especially in resource limited settings which usually have the highest snakebite incidence in the world.

A key finding that can be taken up into an immediate output is an education campaign with two elements. One, to communicate the risk and risk reduction mechanisms for farmers and two, guidance and education on the importance of snake identification. We are going to share our findings through our regional channels and also share with the global snakebite research community, through The Global Health Network [37]. Our study also raised important remaining unknowns; namely, that we need more and better data in a multicentre study to improve on machine learning model performance for predicting patient outcomes.

We undertook this study within a larger research methodology study [38] that is ongoing and seeks to find effective capacity development mechanisms to support care settings, such as SBTRH, to analyse the data they hold to guide their management and treatment practices. Therefore, we shall endeavour to run further studies to address remaining questions and encourage others to do the same.