Ethics

The study was granted ethical approval by the Ghana Health Service Ethics Review Committee (GHS-ERC073/04/24) and followed all ethical guidelines and regulations. All participants were given a clear explanation of the study’s purpose, and written informed consent was obtained. They were informed that participation was voluntary and that they could choose whether to take part in the study without consequences.

Best-worst scaling method

Best-worst scaling (BWS) is a quantitative survey method that typically asks respondents to identify their most and least preferred choices [23], under real/hypothetical conditions, regarding a health program characterized by specific attributes. Participants are presented with various combinations of these attributes across different choice sets, each containing at least three options. A significant advantage of BWS is that it allows respondents to weigh competing characteristics against each other, rather than simply ranking or rating them [24,25]. This approach provides researchers with deeper insights into the relative importance of each attribute. The advantages of BWS, particularly when compared to discrete choice experiments [26], have been noted in a prior study [27]. Results from BWS can effectively inform policymakers about public preferences, which can then be integrated into health program development, ultimately guiding more effective health policies. The use of BWS in health research is widespread [24] and valuable for understanding the factors that influence public preferences for new health initiatives before they are implemented [13,28].

Selection of attributes

The development of the BWS questionnaire was guided by recommended best practices, qualitative research, and reviews of existing evidence to inform attribute selection and design. This process consisted of some key steps. Initially, we created a pool of 18 potential attributes by conducting an online search for features of mobile applications (apps) for snakebite management. This step established a foundational understanding of current mobile app usage in snakebite management. In the second step, we conducted a focus group discussion with a panel of local experts comprising 20 healthcare workers, a public health researcher, and a mobile app developer. The purpose of this discussion was to review the attributes and identify potential issues to enable necessary revisions. The panel provided valuable input, including revisions and the addition of new information, such as ensuring the app offers details on the availability and cost of WHO-approved antivenoms at healthcare facilities and pharmacies. Their feedback helped ensure that the attributes were clear, concise, and easy to understand. Ultimately, we confirmed 11 plausible attributes (functions) for use in developing the BWS questionnaire. The finalized attributes included: Snakebite treatment protocol (app provides intuitive, scientific clinical protocols to follow when dealing with snakebite cases); Step-by-step assistance to victims and first responders (provides clear, actionable clinical and biological advice on steps to take in case of snakebite); Provides education and training (app provides a range of useful education and training materials and links about snake bite envenoming, first aid prevention of snakebite, and snake biology); Identify venomous snake biodiversity in your area/region; Design snake games to teach about snakes in your area/region (provides habitat-specific quizzes); Identify venomous snakes using clinical evidence/symptoms presented by a snakebite victim (neurological, muscular, cardiac, renal effects); Identify venomous snakes in your lab using morphological clues (enlarged teeth; venom fangs generally grooved at the front or with a venom duct orifice near the tip with +/- at the back; there are always normal teeth in front of these venom fangs); Works offline (app designed to run all functions without an active internet connection); Surveillance and monitoring of snakebite victims; Platform for snakebite treatment experience sharing; Information on the availability and cost of WHO-approved antivenoms in healthcare facilities and medicine/pharmacy sales outlets.

Statistical experiment design

To ensure the research design’s effectiveness, we employed a block experimental design with 11 choice scenarios, each replicated twice. Each scenario included five mobile app functions for snakebite management. This method guarantees an equal probability of selecting each attribute in the questionnaire. Since no prior information was available on the parameters, we constructed the design assuming equal choices for each option. Overall, each attribute occurred 5 times across all sets, with each occurring only once in each choice scenario generated from the block design. In this study, each participant was required to make a total of 44 choices (22 for the best option and 22 for the worst).

Study site, participant sampling and recruitment

It is worthwhile mentioning that this study was part of a larger cross-sectional survey conducted from August to December 2024 in two deprived and predominantly rural districts in the Eastern region of Ghana, namely Kwahu Afram Plains North and Kwahu Afram Plains South. These districts are located within the savannah vegetation zone, which includes both the savannah transitional zone and savannah woodland [21,22]. These areas have reported incidents of snakebites [29,30]. Common snakes found in such zones include those with venoms that cause hemorrhagic effects, primarily Echis ocellatus [5]. The clinical manifestations of bites from this snake can lead to renal complications due to coagulopathy, alongside marked local effects such as pain, severe swelling, bruising, blistering, and moderate to severe necrosis. Additionally, Naja nigricollis is known for its neurotoxic effects. Bites from this snake can result in local pain, swelling, blistering, tissue destruction, bleeding, venom-induced coagulopathy, and spontaneous systemic hemorrhaging (in areas such as the gums, gastrointestinal tract, and brain). Thus, healthcare professionals must understand clinical presentations to accurately identify the cause of envenomation and manage the case effectively. Moreover, approximately three-quarters of the communities are located on islands within water bodies, including Volta Lake, the River Afram, and the Obosom River, which limits access to higher-level referral care and leads to delayed referral pathways [31]. Most residents in other settlement communities walk between 25 and 35 minutes to reach the main district hospital in Donkorkrom [31]. Additionally, about 59.9% of the population seeks healthcare outside of their immediate settlement. While information on smartphone ownership among healthcare workers in these districts is scarce, mobile phone ownership among the general population stands at 22.9% in KAPN [21] and 27.1% in KAPS [22]. Since healthcare workers are generally more educated and better connected professionally, it is likely that smartphone access among them is higher. In rural areas of Ghana, smartphones are the most owned functional information communication technology device, with 58.9% of individuals aged 12 years and older using them [16].

The minimum number of respondents required for this study was 70, based on a standard sample size calculation [32] and the available choice scenarios. A 20% upward adjustment was made to account for the expected non-response rate, which added 14 respondents to the total. Consequently, the adjusted sample size for the study was set at 84. However, to enhance the statistical power of our results, a higher sample size was preferable for this cross-sectional study. Therefore, we oversampled and recruited a total of 137 healthcare workers involved in the treatment and management of snakebite patients in the two districts, as similarly reported in our previous study [33]. We employed a multistage sampling technique to select three communities: Donkorkrom, Tease, and Amankwakrom, which served as clusters for our sample [33,34]. From these clusters, we randomly selected 137 healthcare professionals engaged in treating and managing snakebite patients across three health facilities within the communities. The healthcare workers were approached during their working hours at their designated facilities to complete the questionnaires. During data collection, respondents were presented with choice scenarios one at a time. Participants were asked based on their lived experience to select two attributes/functions: one they considered most important and the other they considered least important for the app to provide. They recorded their responses on paper and with a pen. Enumerators ensured neutrality by not observing responses in real time and maintained participants’ autonomy through peer-based data collection. An example choice scenario is presented in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. An example choice scenario.

https://doi.org/10.1371/journal.pntd.0014435.g001

Data management and analysis

Data collection included real-time validation checks to ensure a complete dataset with no missing data [33]. Enumerators conducted initial checks for incomplete responses in the survey tools and encouraged respondents to finish any unfinished sections without observing their responses in real time. No specific issues were identified during the internal consistency checks that could affect our results. To rigorously evaluate ML models performance and reduce the risk of overfitting, we employed a 5-fold cross-validation strategy, which improved model performance compared to the hold-back validation approach. The full dataset was randomly divided into five mutually exclusive folds. In each of the five iterations, four folds were combined to train the model, while the remaining fold was used as a validation set to compute performance metrics. This process was repeated so that each fold served as the validation set once, ensuring that all data contributed to both training and validation and yielding a more stable estimate of model generalizability. In this study, the predictor variables related to mobile app functions used to manage snakebite were treated as generic attributes, while the response variable was effects coded and treated as continuous. Data were analyzed using seven ML models, including Support Vector Machines (SVM), Fit Least Squares, Least Absolute Shrinkage and Selection Operator (LASSO), Ridge Regression, Elastic Net, and Generalized Regression (Forward and Pruned Forward Selection). Model performance was evaluated using the Mean Root Average Squared Error (MRASE), calculated as the average of the Root Average Squared Error across the cross-validation folds, with lower values indicating superior prediction. Additionally, computation time, defined as the time required to fit the model to the training data, was measured to assess model efficiency, particularly for potential real-time or large-scale applications. The corrected Akaike Information Criterion (AICc), and the Bayesian Information Criterion (BIC) are also reported. All statistical analyses for this study were performed using JMP Pro statistical software. The approach employed in the current study, although primarily based on ML models, could not be viewed as solely empirical or entirely dependent on black-box algorithms. By incorporating established methods from the literature, the black box is illuminated, providing insights into the contribution and importance of each attribute in the models [35].

Participant characteristics

This study involved 137 healthcare workers, comprising 83 females (60.6%) and 54 males (39.4%). The participants represented a diverse array of professional backgrounds and areas of expertise, as similarly reported in our previous research [33,34]. Within this participant group, 53 females (63.9%) and 45 males (83.3%) were aged 31 years or older, as reported in Table 1. Importantly, 37.3% of the female participants reported having 3–5 years of experience specifically in managing and treating snakebite patients. This experience level suggests a valuable familiarity with the complexities involved in such cases. The majority of respondents were either certified, enrolled, or general nurses, with 73 females (88.0%) and 23 males (42.6%) falling into these categories. This underscores the critical role that nursing professionals play in the healthcare landscape. Additionally, among male respondents, 12 individuals (22.2%) held positions as physician assistants, clinical officers, or medical doctors, illustrating the range of healthcare roles represented in this study.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Characteristics of the study population by gender.

https://doi.org/10.1371/journal.pntd.0014435.t001

Model performance evaluation metrics

A comparative evaluation of the seven ML models revealed that the Ridge Regression model achieved the lowest MRASE (0.59698), indicating the highest accuracy in predicting the held-out utility scores. The LASSO and Elastic Net models performed similarly well (MRASE = 0.59700). In contrast, the SVR model displayed the highest prediction error (MRASE = 0.62504). In terms of computational efficiency, the Generalized Regression models, specifically Pruned Forward Selection and Forward Selection, were the fastest, with times of 172 milliseconds (ms) and 191 ms, respectively. On the other hand, SVR was the most computationally intensive, taking 9757 ms. Ridge Regression proved to be the most effective models, providing the best balance between predictive accuracy and computational efficiency, with a processing time of 2159 ms. As a result, Ridge Regression was selected as the top-performing model. This model also demonstrated excellent goodness-of-fit, achieving the lowest AICc value of 2787.6966 and the lowest BIC value of 2851.3023. In contrast, the standard/traditional maximum difference model had a significantly higher AICc value of 8240.2924 and a BIC value of 8293.3039. We chose the Ridge Regression model because it offers the best balance between predictive performance, computational efficiency, and overall fit.

Ridge regression model results

The Ridge Regression model was used to assess the relative importance of each mobile app function. The standardized regression coefficients from the final Ridge Regression model were interpreted as preference weights, which were then transformed into utility scores. These utility scores indicate the relative importance of each mobile app function. The results of the full Ridge Regression model regarding the app’s functionalities are provided in Table 2 with an insight based on gender subgroups. The model is statistically significant (p = 0.0001), highlighting the differences in healthcare workers’ preferences for various mobile app functions. Each attribute’s importance was measured relative to the app’s “works offline” function, which served as the reference point for interpretation. However, the functions “identify venomous snakes in your lab using morphological clues,” “snakebite treatment protocol,” “surveillance and monitoring of snakebite victims,” and “antivenom stock mapping” were not statistically significant. This indicates that healthcare workers did not demonstrate a strong preference for these functions compared to the reference function. The analysis indicated that “step-by-step assistance to victims and first responders” was the most important and favored app function among participants, with utility estimate (β) of 0.3924 (95% CI: 0.3183 to 0.4665; p < 0.0001), followed by “provides education and training materials,” (β = 0.2243; 95% CI: 0.1500 to 0.2987; p < 0.0001). This emphasizes the need for practical assistance and educational resources in snakebite scenarios. The attributes “identify venomous snakes using clinical evidence/symptoms” and “identify venomous snake biodiversity in your area/region” were statistically significant and of intermediate importance, reflecting a desire for both diagnostic support and heightened awareness of local snake species. In contrast, “platform for snakebite treatment experience sharing” (β = -0.2503; 95% CI: -0.3268 to -0.1738; p < 0.0001) and “design snake games to teach about snakes” (β = -0.3995; 95% CI: -0.4737 to -0.3252; p < 0.0001) were the least favored or deprioritized features, suggesting a lesser interest in gamified learning or shared experiences in the context of snakebite treatment.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Ridge regression model results for each function of the mobile phone app for managing snakebite envenoming.

https://doi.org/10.1371/journal.pntd.0014435.t002

Interestingly, the subgroup analyses by gender largely aligns with the full model results, though the attribute “identify venomous snake biodiversity in your area/region” is not statistically significant for either male or female participants. “Step-by-step assistance for victims and first responders” is also the most preferred app feature for both genders, with β = 0.4311 (95% CI: 0.3367 to 0.5255; p < 0.0001) for females and β = 0.2928 (95% CI: 0.1734 to 0.4121; p < 0.0001) for males. Following this, “provides education and training materials” had β = 0.2355 (95% CI: 0.1370 to 0.3339; p < 0.0001) for females and β = 0.1820 (95% CI: 0.0688 to 0.2952; p = 0.0016) for males. Conversely, “platform for snakebite treatment experience sharing” and “design snake games to teach about snakes” were significant but were less favored by both gender groups. Additionally, “surveillance and monitoring of snakebite victims” showed significance for females with β = -0.1230 (95% CI: -0.2212 to -0.0248; p = 0.0140).

The differences in each mobile app function compared to the overall average are presented in Table 3. A positive mean difference indicates that a specific function is significantly more preferred than the overall average. Conversely, a negative mean difference suggests that the function is less preferred than the overall average. Among the mobile app functions, those relating to “step-by-step assistance for victims and first responders,” “providing education and training materials,” and “identifying venomous snakes using clinical evidence or symptoms” showed notably higher mean values. In contrast, functions such as “a platform for sharing snakebite treatment experiences,” “designing snake games to educate about snakes,” and “surveillance and monitoring of snakebite victims” exhibited lower mean values. Overall, these findings indicate that healthcare workers prioritized practical, clinically relevant functions that directly enhance emergency response and patient management. Meanwhile, features focused on educational gaming, experience sharing, and surveillance were considered less critical for immediate snakebite care.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Differences of each function of mobile phone app from the overall average.

https://doi.org/10.1371/journal.pntd.0014435.t003