Health interventions that have depended solely on face-to-face delivery can now be widely used outside the clinical settings via mHealth tools. This research indicates that smartphone apps are easily accessible and allow patients to access the required information upon request whenever necessary [31,32]. The main advantages of a smartphone app are to offer multimedia features that combine texts and audiovisual information, comprehensive activation, reproducibility, and final feedback about a health issue [32,33]. However, the limitation of this study is that it used synthetic data for the pilot study, and no human participant data was involved. Additionally, the effectiveness of the app on the target population needs to be validated through a field study.

We developed a cross-platform smartphone app for both iOS and Android users. The app can keep track of work-related risk factors and determine the risk of OUD based on the individual responses. We select several questionnaires from the extensive literature review, and the smartphone app user will respond to those. All responses in the app will be collected and used as an input for the machine learning model to predict the risk of OUD. However, all responses are stored in a cloud-based database that allows for multilayer security via user validation and data encryption. Protecting data from unauthorized access is essential because of participants’ sensitive medical information.

The proposed app was built with the Flutter framework for cross-platform mobile apps. Flutter is an open-source framework by Google for building user-friendly, natively compiled multiplatform applications from a single codebase. The agile development method is used to develop the smartphone app [34]. Software development occurred in 4 cycles. For each cycle, the first step is determining and analyzing the project requirement. Next, the smartphone app is designed with the research team’s input and vision. The research team builds and tests small prototypes of the app to identify flaws. With the feedback, smartphone app adjustments are made accordingly, and the process of analysis, design, coding, and building a prototype is repeated.

In addition, we use Firebase as a cloud-based database, a platform Google developed for creating mobile and web apps. A simplified wireframe of the developed multifeatured smartphone app is described in Figure 1.

The mobile app has several pages, and each page has related functionality. Figure 2 shows app screens and their sequence and functionality.

The following procedure will be used:

An extensive literature review was conducted to identify the risk factors contributing to opioid overuse. A pilot questionnaire consisting of 15 questions has been developed to collect risk factor data. We followed a 3-step process while developing the OUD assessment questionnaire. In the first step, we selected articles containing opioid overuse–related questions that received at least 5 citations. Next, we identified the frequencies of different questions across the selected articles and questions; the questions that had the highest frequency were eventually selected for the project. Finally, our 6-member resource panel evaluated the questions and made small revisions whenever necessary. These questions are classified into 3 categories: 6 of them were highly sensitive, 4 were moderately sensitive, and 5 were mildly sensitive questions. Of the 15 questions, 9 have two response options (yes or no), 5 have five response options, and 1 has three response options. However, all the responses need to convert to numerical values for further processing. Our ultimate goal is to provide a response to the user about OUD based on the response given. A machine learning model NB was used to predict the OUD risk level. All responses were given weights or scores. We have used Likert scales to develop the scoring system, a widely used method in the field of public health. A Likert scale is a validated method to assess attitude and behavior [35,36]. An analysis of various scoring patterns for the risk factors and weights or scores was done on a Likert-type scale. The scoring pattern for the question responses is given in Tables 1-3.

A data set using risk factors and scores was developed. The data set combines various responses and is used as an input for an NB classifier model. Among all machine learning algorithms, NB is used here in this project because it is suitable for binary and multiclass classification. NB works well in cases of categorical input variables compared with numerical variables. NB is useful for making predictions based on past results. Moreover, it is simple. If the conditional independence assumption holds, an NB classifier will converge quicker than discriminative models like logistic regression, so we need less training data. Moreover, even if the NB assumption does not hold, it requires less model training time. The NB model will predict 3 categories: normal, moderate, and high risk of OUD, for the given test data.

This project implements a supervised machine learning model called the NB classifier. The NB classifier is simple and fast, provides satisfactory results for classification tasks, and is easy to implement. NB methods are a set of supervised machine learning algorithms based on applying Bayes’ theorem and using the “naive” assumption that features are independent of each other [1]. Bayes’ theorem states the following relationship between the class variable “y” and feature vector “x₁” through “x_n”:

If we modify all features from equation 1 using “i,” the equation can be simplified to:

which is the probability for a class variable and can be simplified for more class variables to:

which is the probability for each class variable.

We compute the probability of each class of “y” and choose the class with the highest probability as the prediction. NB learners and classifiers can be extremely fast compared with more sophisticated methods. This is one of the reasons to choose the NB algorithm since the app will run on a smartphone, and the app must have a quick response time as a usability feature. The NB algorithm was developed as CategoricalNB by the scikit-learn library [2]. CategoricalNB is one type of algorithm of the NB classifier for categorically distributed data. It assumes that each feature described by the index has its own categorical distribution. For each feature “i” in the training set, CategoricalNB estimates a categorical distribution for each feature “i” of X conditioned on the class “y.” The index set of the samples is defined as J={1,…,m}, with “m” the number of samples.

The probability of category “t” in the feature “i” given class “c” is estimated as:

where is the number of times category t appears in the samples “x_i,” which belong to the class “c_i.” is the number of samples with class c, α is a smoothing parameter, and n_i is the number of available categories of feature i.

The pseudocode for the NB classifier is described in Textbox 1.

We have implemented the machine learning model using the scikit-learn library [2], a free software machine learning library using Python programming. We have used the CategoricalNB algorithm, which implements the NB classifier for categorical features by the scikit-learn library [3] and implements the model on the application powered by Google Colaboratory (Google Colab) TPU with a high-RAM environment, an entirely cloud-based Jupyter Notebook environment. The data set created during feature extraction is used for training the machine learning model. Once the training of the machine learning model is completed, we use the pickle operation to serialize the machine learning algorithm. We create a representational state transfer (REST) application programming interface (API) using Flask, a Python framework used as the back end. We host this REST API onto Heroku, a platform as a service that enables developers to build, run, and operate applications entirely in the cloud. Heroku runs the app in lightweight, isolated Linux containers called “dynos.” We use the free tier for our application. In the REST API, we deserialize the machine learning algorithm back to the Python object and predict the OUD risk responses with input data sent from the mobile app. We send requests to this REST API and questionnaire responses submitted from the mobile app. The predicted OUD risk response will be sent back to the mobile app. The returned response from REST API will be displayed on the Response page. The response from the machine learning model will determine the risk factor of OUD for an individual, whether a respondent is at normal, moderate, or high risk of OUD.

Data have been captured from users for 15 questions, considering the 15 responses as different features for the machine learning model. As we do not have human participants involved in our research, we have created a data set by analyzing the literature review. We have classified these 15 questions into 3 different categories. Six questions are classified as highly sensitive questions, which means there is a high risk of opioid overuse; 4 questions are classified as moderately sensitive, which means there is a moderate risk of opioid overuse. Furthermore, 5 questions are classified as mild sensitive, meaning the low risk of opioid overuse. This data set consists of 15 feature vectors and 1 target vector, which will be the output. Our target vector consists of 3 classes, that is, normal, moderate, and high risk. We have assigned these 3 classes to 3 numerical values (normal=1, moderate=2, and high=3). Each feature vector is assigned a weight or score based on the Likert scale, as mentioned in the Feature Extraction and Processing section. We have created a training data set consisting of 23 samples with 15 feature vectors and a target vector. This data set is used for training the machine learning model.

We have prepared a data set using Likert scale weights, and all the questions or features are considered with equal weightage. We have created a data set by considering weightage to the features based on the category of questions, that is, considering more weightage to highly sensitive questions and less weightage to mild sensitive questions. In order to add weightage to each category of questions, we have considered 50% of the score of highly sensitive questions, 30% of the score of moderately sensitive questions, and 20% of the score of mildly sensitive questions.

As this study did not involve human participants, there were no ethical concerns related to confidentiality, informed consent, or other potential risks associated with research involving human participants. Instead, we used synthetic data generated through simulation methods to investigate the research question. Using synthetic data ensured that there were no privacy concerns or risks associated with identifying individual participants. Therefore, no special ethical considerations were necessary for this study.

A research experiment using a developed mobile app was conducted. The machine learning model has been implemented using an application powered by the Google Colab environment and trained using a created training data set. Once training had been completed, we deployed this model on the Heroku environment and accessed this trained model using REST API from the developed mobile app. We use Firebase as our database for a mobile app that Google is developing. Firebase also offers other services, including authentication, analytics, and so on. Once the user signs up using the required details, the home screen opens, where the user can navigate to the questionnaire screen. The user enters a response to all the questions and submits the responses. Then the machine learning algorithm calculates probabilities for each class of target vector and outputs the class with the highest probability as a prediction. The predicted OUD risk factor will be displayed on the mobile app response screen. Users can also check previously recorded OUD risk factor data in the mobile app. Figures 3-5 show graphs depicting the feature vector relationship with OUD risk response.

OUD is a disease of epidemic proportions in the United States [1]. Studies have indicated that the opioid epidemic has significantly strained the American economy [37,38]. The strains can be felt through the insurmountable health care costs, the strain on the judicial and criminal system, and the accrued loss of productivity [39-41]. The overall outcome of these crises is a deficit of indirect costs associated with increased morbidity and mortality for specific occupations, thus creating a decline in productivity and increasing high costs [41].

Furthermore, individuals employed in areas that have a higher than usual occupational strain—either physical or mental—have been noted to have high premature mortality and loss of productivity, which create an economic burden. Also, occupational injuries because of the nature of the employment result in employees using higher amounts of opioid-based medications [41].

Our study aimed at creating an app used by individuals in high-risk occupations—transportation and health care—to assess individual risk for OUD. An extensive literature review indicated that occupations in which workers are repeatedly exposed to stress and occupations that are physically demanding result in overprescription of opioid-based drugs [18,19]. The study design incorporated mHealth technologies in creating a smartphone app so that individuals in those occupations may confidentially assess their risk level for OUD in privacy at their residence as the traditional questionnaires or inquiries by physicians prove unreliable due to the sensitive nature of the disease. Using an app that can be accessed in privacy ensures an increase in validity [42]. After individuals take the assessment, they can make notes within the app for their provider and are offered resources for getting help. We have created a private pathway for assessing risk and aiding treatment by implementing this technology.

The project has many strengths that offer support in its use. First, the app has been designed for both iOS and Android operating systems, ensuring that individuals can access the app regardless of their preferred operating system. The app is user-friendly and consists of 15 questions. This feature was created after extensively reviewing the literature to determine risk factors and behaviors highly correlated with OUD. The questions are categorized according to high-, medium-, and low-risk factors ensuring precision and the risk analysis for OUD at an individual level. Furthermore, studies indicate that health care apps are highly effective in shaping health attitudes, empowering patients, and “pushing” them in the direction of managing their condition appropriately [43-45]. This ultimately lends itself to the app having a precise related functionality.

The app’s most highlighted strength is that it uses an NB model in statistically analyzing the 15 questions individually. The NB model has been widely used for classifying and clustering outcomes, but this project uses the model and probabilistic interface [46]. Machine learning models, such as NB, have been used exclusively for data mining. However, the current trend is to use them in disease prevention [47]. The app also uses a REST API, thus further ensuring the confidentiality of the patient.

As with all apps, there are some limitations. Self-administered instruments run the risk in that the respondents do not answer truthfully or minimize their responses to ensure a desirable outcome or appearance of the responses [48,49]. Currently, the app has undergone several pretesting phases to assess the accuracy of analyzing the respondents’ responses. We are still in the process of working with Apple to launch the app commercially. As a result, we cannot guarantee that the graphical user interface is navigable for our target audience [50,51].

The future goals of this project are to expand the app’s functionality and include some educational intervention features. The proposed app will be used as a launching pad for various mHealth-based assessments. Moreover, we used an artificial intelligence–based algorithm, and as the number of users increases, the algorithm can predict better risk factors. After conducting several tests or pilot studies of the app, we can administer this app to the communities identified as high-risk, such as health care, construction and extraction, transportation, and warehouse industries.

The current trend indicates that use of mHealth techniques, such as a mobile app, is favorable in predicting and offering mitigation plans for disease detection and prevention [47,52-55]. Amidst the ever-increasing opioid epidemic, occupations of note, such as transportation and construction, have encountered alarming issues related to opioids. Because of the intense physical and mental strain and stress placed on employees in these fields, studies indicate that they face a higher risk of developing OUD [18,19].

Recognizing the current movement to use mHealth technologies and the sensitive nature of self-reporting one’s level of risk for developing OUD, the development of our smartphone app offers a promising tool for addressing the opioid crisis in specific workforces. The use of artificial intelligence in predicting an individual’s OUD risk offers a novel approach to prevention and intervention. The implementation of mHealth tools, such as our app, has the potential to provide convenient and effective interventions outside of traditional clinical settings. We hope that the development of this app will contribute to a reduction in opioid misuse and provide tailored mitigation strategies to communities most impacted by OUD. Further research is needed to validate the effectiveness of our app and to ensure that it meets the needs of the target population.