The eligibility criteria were developed through collaboration between authors AB and JM. Papers included in this review carried out at least one action related to machine learning and medication adherence. Studies also had to test out their application of machine learning to medication adherence using either real patients, research participants, or simulations. The ABC taxonomy was used to define medication adherence for this study. This taxonomy defines medication adherence as “the process by which patients take their medications as prescribed, composed of initiation, implementation and discontinuation” [4].

The following databases were searched to identify relevant papers: PubMed, Scopus, IEEE, ACM Digital Library, and Web of Science. These databases were selected by the Laboratory of Applied Informatics Research (LAIR) team as representing the largest and most extensive coverage of studies that investigate applying machine learning to medication adherence.

Search queries for this paper were created with the help of Rebecca Carlson and Fei Yu, both of whom are experienced research librarians. Search terms were selected by reviewing exhaustive term lists provided by the research librarians on the topics of machine learning and medication adherence. The same keywords were selected for all five databases by combining the lists of machine learning and medication adherence terms using an AND operator. The final search was conducted on April 30, 2020. When searching PubMed, Medical Subject Headings (MeSH) terms were incorporated to provide broad coverage of related vocabulary. Following PubMed, this same query strategy was used to search Scopus but with Emtree Terms in place of the MeSH terms. For IEEE, ACM Digital Library, and Web of Science, the keywords were used exclusively to discover relevant papers. Examples of the search queries are provided in section A of Multimedia Appendix 1.

The search queries found 504 studies that contain both machine learning and medication adherence terms. After removing duplicate studies, 417 papers were identified for review. The title and abstract review were conducted using a two-person team consisting of authors AB and MK. The title and abstract evaluation reduced the total number of relevant papers to 54. Next, a sample of 20 articles was screened. This initial screening was done to ensure that information was collected in a consistent manner using a data charting template. This template was developed through collaboration with the LAIR team and the primary author of this work. After finalization of the data charting template, all 43 papers were reviewed in full by AB and MK. Studies that passed the full-text review focused on both medication adherence and machine learning. A focus on medication adherence and machine learning was essential to reduce the number of studies to a manageable list of highly relevant works. Included works also evaluated the use of machine learning for medication adherence actions using either patients, research participants, or simulations. This testing requirement was used to exclude studies that are still in the early development stages and to make sure that included works contained all of the relevant data charting elements. Next, the results were grouped into natural categories and analyzed according to the combination of medication adherence actions in each study. This grouping of related studies was carried out by AB, JM, and MK.

Figure 1 shows the article review process and provides reasons why specific studies were excluded during the full-text review.

A data charting form was created and updated throughout the review process as needed. This form was used to guide the selection of relevant information throughout the review process and was tested to ensure consistency across different articles during the sample screening. The items included in the data charting template are listed in the next section.

Data charted for this review includes the article title, publication date, study number, study goals, main study results, disease states, predictors of medication adherence, types of machine learning used, number of participants, data collection methods, actions related to medication adherence, adherence measurements, limitations, and inclusion or exclusion status.

Following data charting, actions related to medication adherence were determined for each study. Next, articles were grouped for further analysis according to the combinations of medication adherence actions included in each work. AB, JM, and MK conducted the categorization process, and disagreements were discussed until a consensus was reached. Following categorization, the data charting documents for these works were reviewed to find relevant themes. A general analysis was also conducted before categorization in terms of the number of publications per year, articles per disease state, and publications per database type.

This paper includes the analysis of 43 studies selected for inclusion. The main results of this review are listed in Tables 1-3. More detailed data charting tables are also available in section B of Multimedia Appendix 1.

Before dividing the studies into categories, they were examined as a whole to determine the distribution of articles concerning time, disease states, and database type.

Table 4 displays the number of publications per year that apply machine learning to medication adherence. For the past few years, this topic has generated a growing number of publications, indicating that there is interest building around this topic.

Table 5 illustrates the distribution of different disease states within the included articles. The largest group in this graph is general and consists of 12 studies. General indicates that these studies did not focus on any specific disease or group of diseases. The remaining 32 papers, except a single study focused on high-cost medications, looked at chronic diseases.

Following the general analysis, actions related to medication adherence were determined. The three identified actions were prediction of adherence, adherence monitoring, and adherence interventions. Next, studies were grouped into natural categories for further analysis according to the combination of medication adherence actions that they contain. The following natural groups were identified: prediction of adherence only, monitoring of adherence only, monitoring with an intervention to improve adherence, all three medication adherence actions simultaneously, and prediction with monitoring.

The fourth group contained 3 articles that predicted adherence, monitored adherence, and intervened to improve medication adherence [43-45].

The first of these studies allowed patients to request medication refills using a conversational AI SMS text messaging system [43]. The system would prompt the user when they needed to request a refill for their medication. The predictors used within this study were age, gender, spoken language, address (used to determine social determinants of health), race, ethnicity, and patient response via either structured response or using free-text entry. This study used an artificial neural network to predict if the patient would use this system to request a medication refill. Machine learning was also used for conversational AI, but the specifics were not discussed within the article. This system was not restricted to any specific disease state and was tested using approximately 99,000 patients. The limitations are that data was only collected over 3 months and that 8% of free-text responses were not correctly understood by the conversational AI.

The next work used data from the 99DOTS (Directly Observed Treatment Short Course) program, which monitors medication adherence of patients with tuberculosis [44]. Patients in this program used a phone to enter a random code for each dose of the medication they took. Other data collected for 99DOTS included: demographic data (age, weight, gender, treatment center ID), treatment start date, treatment end date, adherence string, and treatment outcome if applicable. This data was manipulated to create both static and time series predictors. Random forest, linear regression, and support vector machine (SVM) were applied to the static predictors to determine daily nonadherence risk, treatment success, and how to best allocate limited resources. A deep network called LEAP, which is capable of using time series predictors, was also used to predict these same outcomes. LEAP had the best prediction accuracy of the machine learning methods used. This system was tested using the monitoring data of about 17,000 people.

The next paper used face recognition software and computer vision to monitor medication adherence of 53 patients that had schizophrenia [45]. Patients in this study used a smartphone camera to record medication ingestion and submitted those recordings to the research team. Face recognition and computer vision techniques were then used to flag patients engaging in suspicious behavior, which indicated that long-term adherence was unlikely. This system also reminded patients to take their medications at a specific time each day. The main limitation was that patients were allowed to choose either direct in-person observation or monitoring via the app thus introducing bias into the study.

The next group had 2 articles that predicted and monitored medication adherence but did not introduce any medication adherence intervention [46,47].

The first of these articles used data collected during hospital stays to generate predictors. Of these predictors, disease severity and biomarkers (breath, saliva, blood) had the largest impact on the model’s accuracy [46]. The overall goals of this study were to predict the adherence risk of a given patient with heart failure and to monitor their ongoing adherence. Nine different classifiers were tested for this study, including random forest, logistic model trees, J48, simple classification/regression tree-CART, rotation forest, radial basis function network, SVM, Bayesian network, and Naive Bayes. Of all of these methods, random forest performed the best in terms of classification accuracy. The data of 29 patients was used to create this system.

The second paper used interactive voice response assessments to predict future adherence and to monitor current medication adherence in patients with depression [47]. The data used for this study was collected using interactive voice response assessments of 208 patients. These assessments were then used to create predictors. Of these predictors, the following added to the power of the model: past medication adherence, age, and physical functioning at baseline. Only logistic regression was used to predict future medication adherence.

This is the first review to focus broadly on applying machine learning to medication adherence. This study provides a general summary of the topic and categorizes literature according to the combination of medication adherence actions. Within each category, common themes were explored and opportunities for future work were identified.

The application of machine learning to medication adherence is a topic still in its infancy that has become more prevalent over the last few years. This technology is typically being applied to patients with chronic diseases that require long-term medication use. In fact, 29 of the 43 studies looked at using machine learning to impact medication adherence within the context of chronic disease [6,7,11-15,17,18,20,23,24,26, 29-33,35,37-40,42-47]. This is not surprising, since patients with chronic diseases often have a high medication burden and poor medication adherence over an extended period of time.

Twenty predictors of medication adherence were found to be important across two or more independent studies [6-19]. These predictors included: education level, marital status, income, gender, geographic location, emergency care interventions, age, race, ethnicity, disease severity, comorbidities, medication cost, insurance coverage, substance abuse, medication beliefs, medication knowledge, medication dose, medication frequency, initial medication adherence, and current medications. This suggests that it may be possible to build models that maintain a high level of accuracy even when applied to a different set of patients. In fact, 1 study in this review applied an established algorithm developed using the MEPS data set to a new data set pulled from a social health forum. This study managed to achieve a modest prediction accuracy of 67.8%, despite being trained on a completely different data set. However, this required careful manipulation of the input data to ensure it was formatted appropriately for the established prediction algorithm. In addition, two predictor variables were discarded to improve prediction accuracy. One limitation of several studies that predicted medication adherence is that they did not take socioeconomic factors into consideration [16,18]. Socioeconomic factors were determined to be strong predictors of adherence by multiple studies within this review and should be included when possible [7-9,17].

Similar methods of data collection were also used in works that solely predicted medication adherence. The most popular method of data collection and determining adherence was self-reported questionnaires. In fact, 8 of the 15 studies in this group determined medication adherence using self-reported questionnaires [5-7,9,11,12,15,16]. Although collecting data with questionnaires is attractive due to its simplicity and low cost, it typically overestimates adherence when compared to other data collection methods. When a questionnaire is used, a validated self-adherence measure is preferred, and efforts should be made to reduce negative social pressures associated with reporting nonadherence [48]. Pharmacy claims data was also used by one-third of these studies for both patient information and to determine adherence [8,10,13,17,18]. Pharmacy claims data is generally considered more reliable than self-reported questionnaires but still may overestimate adherence since it does not actively investigate whether the patient is taking the medication or not.

Additionally, studies that only predicted medication adherence used many of the same algorithms. Of 15 studies, 13 in this group used either logistic regression, artificial neural network, SVM, or random forest algorithms. Some of these works compared the different types of algorithms to determine which was the most accurate [7,12,14,15]. However, more publications need to be generated before any type of consensus can be reached on this issue.

Furthermore, 13 of the 15 studies that attempted to predict medication adherence used either a self-reported adherence metric or medication filling data to train their models [5-13,15-18]. This adherence metric along with predictor variables was provided during the training of each algorithm. Following model training, the adherence metric was not provided to the model and instead was predicted based only on the predictor variables. Of the 6 studies that provided prediction accuracy, 5 of them reported accuracy of 77.6% or higher [5-7,12,13]. This level of accuracy is impressive, and this data could potentially be leveraged to prevent medication adherence issues before they occur. However, the specificity of the algorithm in question would need to be carefully investigated to ensure that it is not overpredicting poor medication adherence, potentially wasting resources directed at preventative measures. Within this group, 5 other studies reported a statistically significant association between predictor variables and medication adherence as their main result [8-10,17,19]. One study found that both nationality and provider type are significant predictors of medication adherence [10]. These two predictors should be investigated in follow-up studies to determine if they remain significant across multiple investigations. Another study within this review found a significant association between specific keywords and medication adherence, which can be used to predict medication adherence based on free-text discharge summaries [19]. More studies should be conducted to find keywords associated with medication adherence since a large amount of medical documentation is stored as free text. This is especially useful in analyzing data from less sophisticated clinical systems with less structured documentation.

Machine learning was also used to monitor medication adherence. The main purpose of all 13 studies within this group was to develop new ways of monitoring medication adherence with the aid of sensors and machine learning. One popular approach, used in 6 of these 13 studies, was to attach sensors to the container holding the patient’s medication [22,23,26,28,30,32]. Four of these studies attached a microphone to an inhaler, and the other 2 used a smart pill bottle to track medication adherence. Direct attachment of sensors to medication holding devices is an established method of tracking medication adherence since the opening or activation of the medication holding device can be directly recorded. However, this form of data collection still has some major drawbacks outside of smaller deployments because it requires copious hardware to scale. This hardware would also likely be disposable unless additional effort was required on the part of the patient to load the medication or medication container into a reusable device.

Another popular approach was to use portable smart devices to track medication adherence. Of these studies, 6 used portable sensors found in common consumer electronics like smartwatches or smartphones [20,21,25,27,29,31]. One of these studies required users to actively record ingestion using a smartphone camera, and the other 5 recorded movement data using smartwatch sensors. Having patients track their medication ingestion with a camera every time they take a dose is accurate since the act of ingestion is being directly observed. However, it is unclear if patients would actually use the system and might find this level of observation overly intrusive.

Studies that only monitored medication adherence used many of the same machine learning algorithms. Of the 13 studies, 11 in this group used logistic regression, artificial neural network, SVM, or random forest algorithms [21-26,28-32]. Several studies compared the accuracy of these different algorithms, but this was not enough to generate any meaningful conclusions [22,26,28,30].

Moreover, works that only monitored medication adherence often built systems that classified user activities based on data provided by sensors. Several studies attempted to monitor the use of inhalers by analyzing audio recordings of inhaler use [23,26,30,31]. All 4 of these studies were able to classify sounds of inhaler use with an accuracy of 93.75% or higher. This is especially impressive when one considers that 3 of the studies also introduced background noise samples to make the evaluation more comparable to real-world environments [26,30,31]. All four systems also classified different parts of inhaler use such as inhalation, actuation of the inhaler, and exhalation. This allows these systems to track both medication adherence and proper inhaler technique simultaneously. Patients using a device like this could be coached in real time using an app or could be flagged for inhaler use training by a health care professional at a later time. In addition, since these systems are only using audio samples, similar results could probably be obtained using a smartwatch with a microphone.

Two other studies observed the movement patterns of patients with Parkinson disease [24,29]. Both of these systems were able to classify the medication status of the patient with an accuracy of 78% or higher, one using body worn sensors and the other using a stationary sensor mounted in a room. This level of accuracy is impressive and has potential that goes beyond simple medication adherence. Systems like this could be used to help inform clinicians of when a patient’s medications need to be adjusted. This is helpful for patients in more advanced stages of degenerative diseases who may have difficulty communicating their current status to care providers. Three systems also classified medication administration using only movement sensors in a smartwatch. These studies were able to achieve a classification accuracy of 78.6%, which is impressive, but this was in a controlled setting with the number of possible activities restricted to a small list [20,21,25]. Determination of medication administration using only movement sensors in a smartphone is an attractive option since it allows tracking with devices that many patients are already using. However, this technology still requires more testing in realistic environments before it can be trusted to accurately determine administration.

A total of 10 studies also monitored medication adherence and introduced an intervention to improve it [33-42]. Some of the works within this group used many sensors in conjunction with alerts to provide context-aware reminders [33-36]. Two systems using context-aware reminder systems were able to keep patients at an adherence level of 92% or higher [33,34]. This is an impressive level of adherence, but these systems were intrusive and required the use of a large number of sensors throughout the living spaces of the user. Additionally, only 1 of these studies compared its results against a traditional reminder system, and they did so with the results of only a single patient [34]. Furthermore, none of these studies mentioned any effort to establish if users were comfortable with the level of monitoring required to produce a context-aware reminder system. This is potentially an oversight since these systems require the use of multiple sensors throughout a person’s home. The companies creating smart devices would also need to be open to data sharing with their competitors if systems like this are going to be possible in the future. Other works that included medication monitoring with an intervention used conversational AI to communicate with patients [38-40]. All of the conversational systems included reminders and were able to interpret free-text responses provided by patients. Additionally, all of the conversational AI systems were compared against a control group to show if they improved adherence. Two of the three conversational AI systems significantly improved adherence over more traditional reminder systems. One of the conversational AI systems improved refill rates by 14.07% and the other improved average adherence by more than 20% over the study’s duration [39,40]. However, the third conversational AI system did not significantly improve medication adherence when compared to a more traditional system [38]. In general, conversational AI systems are further along in development, allowing the authors to focus more on comparing their systems to more traditional solutions.

The analysis of this topic was limited to five important databases, and some relevant articles from other sources may have been missed. Different grouping strategies might have also added additional insights but were not attempted since they are outside the scope of this paper.