Owing to the large amount of information available in health institution databases, including medical treatments, diagnostic tests, clinical histories, and drug characteristics, there is a need to implement recommender systems (RSs) that support medical staff in activities related to health control and management. The main concept of an RS is to suggest items that are particularly suitable for the user based on their profile or historical preferences. In the context of health, these items can be drugs, medical treatments, health videos, and patients sharing the same disease. An RS employs data sources to learn about user preferences through machine-learning algorithms and information-filtering techniques such as content-based, collaborative filtering–based, demographic, and hybrid approaches [1].

Currently, more than 80% of internet users seek health information through various platforms, including social networks, a figure that, according to De Choudhur et al [2], continues to grow. Through the internet, users can identify patients with the same disease, the possible causes of their ailments, find procedures to alleviate a particular disease, learn new healthy habits, and find general health information [3-5].

In the case of chronic diseases such as diabetes, the prescription of multiple drugs is common; thus, RSs can support the intervention of the treating doctor in determining which drugs to prescribe to a particular patient. Considering the current health status of a patient, their clinical history, medications prescribed in previous periods, specific symptoms, and other characteristics, the system can search for patients with similar parameters in the database and suggest drugs that have been more successful in these cases, which could be recommended to the target patient.

Some research has been performed on RSs in the health area. For example, Zhang et al [6] proposed the iDoctor system to provide users with personalized medical recommendations. This system explores users’ emotions and preferences about doctors through their ratings and reviews. Gujar et al [7] proposed data-mining techniques for the prediction of a disease and to make a recommendation for specialists about the predicted disease. Kuanr et al [8] proposed an RS for cervical cancer in which predictive models showed high accuracy. Poornima [9] presented a daily nutrition RS for women taking into consideration physical data, preferences, and personal information to combat diseases such as malnutrition, obesity, and cardiovascular diseases.

Other research has focused primarily on recommending doctors and hospitals that are best suited to a specific patient profile [10], medication recommendations [11], treatment recommendations for patients over time [12], videos about health [13], and even customized meal plans [14].

Recently, several studies related to the use of RSs in diabetes have emerged, including some exploratory analyses on the disease, predictions on diet plans to combat obesity and diabetes [14], and physical activity and diet plan RSs to help prevent chronic diseases [15].

Clustering is one of the most widely used machine-learning techniques in the field of health to identify patterns or groups of patients with similar characteristics [16,17]. Although the clustering technique has been the subject of research in the area of RSs, these systems have not yet been widely used in medicine. Moreover, technologies that enable only the analysis or prediction of diagnoses or diseases would not be sufficient to provide personalized care to the patient or to support health personnel in making decisions about which drugs to consider for certain diseases. Therefore, we here propose a drug RS for patients with diabetes based on clustering techniques as a complement to the treatments given by the treating doctor.

Both drug predictions and recommendations were evaluated using traditional metrics to measure the performance of the RS.

The aim of this study was to complement previous diabetes-related studies by first analyzing data related to patients with diabetes to obtain important information for the management of this disease, followed by identifying groups of patients who share similar characteristics, which could enable discovering patterns of interest that can support decision-making. Finally, an RS was developed that suggests medications for diabetes according to the patient’s historical information and the doses of the medications administered.

The data set used was obtained from the University of California Irvine (UCI) Machine Learning Repository [18], which contains information on patients with diseases associated with diabetes [19]. The original data set includes more than 50 features representing patient outcomes from 130 US hospitals. This data set has more than 100,000 patient records, which refer to 10 years of health care records (from 1999 to 2008).

In summary, the information contained in the data set refers to: admissions of patients to the hospital; information on 24 medications administered to patients with diabetes; changes in the patients’ medication, and whether the dosage was increased, reduced, kept stable, or not administered; number of medications administered to the patient; time spent by the patient in the hospital, recorded in days; results of tests that were indicated to patients prior to and during their treatment; diagnosis; type of admission; specialty of the treating doctor; and patient data such as age, race, and gender.

Exploratory analysis of the data set is a critical process in research to discover patterns, detect anomalies, test hypotheses, and test assumptions with the help of statistics and graphical representations. It is good practice to first understand the data to obtain as much information as possible.

Toward this end, in the initial analysis of the patient data, we determined that most of the patients belong to the age range of 50 to 80 years, and we classified patients according to readmission status (ie, if the patient presents a case of readmission greater than or less than 30 days, and no readmission). Figure 1 shows the relationship between age and patient readmission, demonstrating very few cases of readmission for younger patients (under 40 years of age). In addition, women had a slightly higher readmission rate than men in cases of readmission longer than 30 days. Readmission showed a similar distribution for patients with and without medication prescribed for diabetes prior to hospital treatment. In addition, we determined that the majority of the patients were of the Caucasian race and did not have a glucose or hemoglobin A1C test.

First, PCA was used to reduce the dimensionality of the data set, resulting in 8 components explaining most of the variance of our data (ie, these components explained 62% of the total variance). We then applied the two clustering algorithms with the 8 principal components. Table 5 shows the best results of the experimentation with the K-means and DBSCAN algorithms.

In comparison with DBSCAN, the K-means algorithm had a higher Silhouette coefficient and a much lower number of clusters for the same data set. Moreover, K-means had a much faster execution time, which means that this algorithm presents lower computational complexity in terms of execution compared with that of DBSCAN. Therefore, the clustering results obtained with K-means were further considered for the calculation of prediction and recommendations. Analysis of the clusters was performed considering the clustering obtained with the K-means algorithm, which showed the best results.

Cluster 4 had the highest number of patients, followed by clusters 2 and 6. Cluster 3 had the smallest number of patients due to the different characteristics considered to group similar patients, such as the diagnosed diseases, race, main drugs administered, and most representative age range.

Figure 3 provides details about the six clusters, including the variables analyzed, such as age, gender, race, health problems, medications, and information about the doses of insulin administered to the patient.

To evaluate the quality of the predictions, we used the mean squared error (MSE), which penalizes more severely when the error is higher [20]. A significant MSE value of 0.53 was obtained in the test set, which means that the system is capable of obtaining good predictions. As a measure of error, a lower the error value indicates a more efficient RS [21].

The quality metric is fundamental to measuring the performance of our RS, since it provides information on the proportion of recommended drugs that are relevant for the user. In our case, a drug is considered relevant when the value of the dose is greater than 1. In the experiments, a precision value of 0.61 was obtained, which indicated that the system provides acceptable recommendations.

Considering a patient “p” with diabetes who belongs to cluster “c,” a list of drugs {f₁,f₂,...f_n} with the highest prediction score is recommended by the system. An example of the recommendation for two patients in cluster 1 is presented in Table 6 based on a setting of providing the top 3 recommendations.

Table 6 shows that the recommended medications were similar for both patients in cluster 1; however, the order of recommendation varied according to the prediction score obtained for each medication. This is to be expected since these two patients share similar characteristics in terms of clinical information, personal data, and medical treatments stored in the data set.

Our analysis of the UCI Machine Learning Repository showed that most patients with diabetes have circulatory and respiratory problems, followed by metabolic and nervous system problems. Regarding gender, women with diabetes showed more circulatory and respiratory health problems compared to men. It was determined that diabetes manifests differently for individual patients. In addition, in the analysis according to race (Caucasian, African American, Hispanic, and Asian), different categories were identified according to patient and clinical characteristics: (1) older patients with circulatory, respiratory, and other problems; (2) younger patients with digestive, respiratory, and other problems; (3) patients requiring increased insulin doses; and (4) patients prescribed more than one type of medication.

From the data set used, we further observed that diabetes can occur at any age; however, the disease appears to be more common among middle-aged and older people. The analysis showed that the health outcomes related to diabetes are different for each patient, both at the level of control with medication and health complications. Based on these findings, an RS is required to provide support to health care professionals to facilitate the management and control of this chronic disease. Therefore, a cluster-based RS was proposed to help recommend drugs to patients with diabetes. The clustering technique was used to identify groups of patients based on their similar characteristics, and the collaborative filtering technique was used to present information on the doses of the medications administered to patients. Our experimental results showed acceptable performance of the proposed system.

The use of explicit information (ratings) from users enables making more precise recommendations; however, according to Wasid and Ali [22], additional effort is required from users when rating an item. Therefore, obtaining useful ratings without additional effort from users is one of the challenges of RSs based on collaborative filtering.

A typical problem of RSs based on collaborative filtering is that data and ratings are often scarce, resulting in a problem of new user and new item cold start. An alternative solution to overcome this problem is to calculate the similarity of the users based on user profiles [13]. For example, if two users are diagnosed with the same disease, they could be considered similar, even if they have not been administered the same drugs during their treatment. Other characteristics such as gender, age, medications administered to the patient, and diseases diagnosed could help classify patients into clusters. Although this study avoided the cold-start (new user) problem by using patient information for clustering and subsequent recommendation, there is a limitation with medications that are entered in the system and have not yet been prescribed to any patient, since the system would have difficulty recommending such medications (ie, new item cold-start problem).

One solution to this problem is to use the metadata of the new items when making recommendations [13]. Combining both user and item information would provide a hybrid approach to address the new user and new item cold-start problem. Therefore, as future work, we propose to (1) extend the recommendation approach using drug information for the prediction and recommendation process, and (2) consider the clustering results of the DBSCAN algorithm for prediction and analyze whether this can improve the quality of recommendations.

We found some previous studies related to the topic of RSs in the health domain; however, our proposed approach differs from these previous works by focusing on combining collaborative filtering with clustering techniques to avoid the cold-start (new user) problem.

The experimental results showed that our recommendation approach performs well in terms of offering good predictions and acceptable recommendations considering patient information. In comparison with the study of Sanchez et al [13] who recommended educational content about diabetes, our work focuses on recommending medications to patients with diabetes. Galiano and Paccanaro [23] used collaborative filtering for the prediction of medication side effects to provide recommendations to safety professionals based on a latent factor model. A latent factor method could also be considered for drug prediction to patients with diabetes, and these results can be compared with those obtained using the clustering-based recommendation approach. Consequently, RSs have been developed based on multiple methods that could be combined with clustering techniques to improve the recommendation process. For example, Chung and Jung [24] proposed a knowledge-based cluster model to improve prediction accuracy and make health care recommendations.

We present an RS that is capable of suggesting medications suitable for patients with diabetes. This system considers user metadata to alleviate the cold-start (new user) problem, obtaining groups of patients with similar characteristics using clustering techniques, which are then used to recommend drugs for patients in the same cluster.

To measure the performance of the recommender system, the quality of the predictions and recommendations was evaluated. In the case of prediction accuracy, a significant MSE value was obtained and acceptable accuracy was found in the quality of recommendations, which can be further improved by using information from more drugs or combining with another collaborative filtering approach such as an item-based approach.

The proposed system offers a new method to provide support to health care personnel during the medical care of patients with diabetes by offering recommendations of possible medications that can be considered for the treatment of this disease. In addition, our RS has the advantage of providing recommendations that can be easily explained since the system recommends drugs that have been administered to patients with similar characteristics to the target patient. We believe that this system could be an important tool for health personnel, as it would help to streamline the process of health care and management.