The development of new technologies shows promise for causing a revolution in the way chronic diseases are followed and treated [1]. In the past few decades, we have seen the following two major technological changes directly connected to the availability of medical information: (1) introduction of electronic health records in most health care facilities, and (2) accessibility to portable devices capable of acquiring information about their users. Both systems are already being used to enhance communication between health providers and final users and to improve the overall performance of health care. Indeed, public and private entities are massively investing in the development of web-based platforms or smartphone apps through which patients can organize their medical agenda, have access to all or part of their medical records, provide their input, and join their medical referents [2].

It seems reasonable to believe that the follow-up of persons with mental illness will be improved if eHealth systems lead to an increased interaction with health care providers. Our group and others have shown that electronic assessment is feasible with proper adaptation [3]. Efficient monitoring may prompt health responses in cases of emergency [4], inform accurately about real-life behaviors between medical appointments, reduce unnecessary visits, and sustain therapeutic decisions [5]. Other parameters, such as biomarkers and input from close relatives, can be added to the monitoring system. This kind of ecosystem already exists [6], and a growing body of evidence has shown that eHealth tools improve treatment outcomes in terms of engagement, symptom improvement, well-being, and self-care [7-13]. Their combination with machine learning techniques has also shown positive impacts on the diagnosis, prediction, and prevention of several diseases, such as cancer [14-16].

Despite these advances, the majority of potential users, such as elderly persons with low educational levels [17], seem to be unenthusiastic about e-mental health tools [18]. Utilization rates have been associated with the characteristics of health professionals [19], but there is little knowledge about the kinds of patients who use e-mental health, how they become users, and what are their patterns of use. Young age, high education, and dissatisfaction with the health care system might be common features among eHealth users [20]. We do not know if this profile also applies to patients with mental disorders, but their digital phenotype [21] is likely to contain valuable information for clinicians and providers alike [22,23]. Alterations in the patterns of use could help clinicians to detect pathological or risky behaviors and individual needs, and increase treatment efficiency.

In this article, a nonparametric latent feature model based on the Indian Buffet Process (IBP) explores the response patterns of 2254 psychiatric outpatients to a web-based questionnaire. We aimed to define specific profiles according to response patterns and then link profiles with psychiatric disorders. The study will thus specifically describe patients with psychiatric diagnoses who have used an eHealth application at least once. We hypothesize that data mining techniques, such as IBP, can be used to associate information from different questionnaires and assessments in a plausible model that could serve ultimately to plan health care delivery.

The data modeling approach we applied was able to discriminate four different profiles of patients based on the answers to a brief electronic questionnaire. All profiles shared a component associated with feelings of aggressiveness, worthlessness, and suicidal thoughts (bias term or profile 0), which seemed to be common among patients who used e-mental health tools [33], such as the MEmind Wellness Tracker.

Sex and age distributions showed very little variability across the profiles, facilitating comparisons between them. In addition to profile 0 (bias term, default pattern), three profiles where found based on the scores of different sets of questions. It is important to bear in mind that a factor set was classified as active only when the scores were in the top 10% of the corresponding items. For example, even if the item of low sleep quality is absent from profile 2, a patient with that profile could still have high scores in that item compared with the general population and thus have relatively low sleep quality.

Patients in profile 1 reported a lack of positive mood, low quality of sleep, low energy, feelings of loss of control, and difficulties to face problems. These symptoms could be reactive to life difficulties and partly due to a lack of coping skills. Patients in profile 2 presented high scores in depressive feelings, worries interfering with their sleep, feelings of being overwhelmed and unable to overcome troubles, low self-confidence, and feelings of worthlessness. This pattern seems to be related with a greater inward focus and depressive-like symptomatology. Interestingly, patients in profile 2 also reported the highest scores for suicidal thoughts and feelings of aggressiveness. Indeed, patients in profile 2 reported five of the 10 ICD-10 diagnostic criteria for a depressive episode, including disturbed sleep, depressive feelings, reduced self-confidence, ideas of worthlessness, and ideas of suicide [26]. Surprisingly, those in profile 2 were evaluated by their physicians as having a higher level of functionality (CGI) than those in profile 1, despite higher levels of suicidal thoughts, aggressiveness, and depression in profile 2. This points to discordances between the medical assessment and the self-reported momentary assessment. Finally, profile 3 involved a small group of patients with high scores in all the items of the questionnaire. They shared the features of profile 1 and profile 2, and reported the most severely affected psychological state in our sample (the highest levels of distress).

Our study suggests that the analysis of data from electronic self-assessments can discriminate profiles or clusters of patients sharing similar clinical characteristics. These features do not seem to overlap with usual clinical diagnoses, since no differences were found in the prevalence of previous psychiatric diagnoses between profiles. Most patients in each profile received diagnoses in F4 (anxiety disorders) and F3 (mood disorders) ICD-10 categories, which were numerically the most common diagnoses in the sample. However, diagnoses of disorders with an onset during childhood and adolescence (eg, F8 and F9) and schizophrenia (F2) were overrepresented in profile 2. Profile 3 was particularly overrepresented among the small group of patients with organic mental disorders (F0) in the sample, which could implicate a more complex disease course. Interestingly, in a previous paper, we found that the assessments made by clinicians did not correlate well with patients’ self-reports within 24 h of a clinical evaluation [34].

The presence of sporadic suicide thoughts can be relatively frequent in psychiatric patients, but eHealth apps could help to identify profiles with higher suicide risk, such as profile 2. Previous literature has suggested improvements in mood, well-being, anxiety, and self-awareness, as well as a higher adherence to treatments among users of eHealth apps [5,35-37]. Electronic assessment tools, such as the one used in our study, may support physicians to discriminate patients with high suicide risk in order to adjust their interventions.

Among the limitations of our study, we note the use of only baseline assessments and incomplete clinical information. The described profiles might not be reflective of eHealth users who continue to use the app regularly. Besides, our intention was not to map the participants onto Diagnostic and Statistical Manual of Mental Disorders (DSM) or ICD categories but rather to identify symptomatic profiles that are not necessarily reflected in psychiatric diagnoses. This study was designed to explore the utility of a new method to classify e-mental health users, and it needs to be completed with follow-up data. Nonetheless, once the SPFM is trained, it will be possible to analyze changes in patient profiles during continuous assessment with several time points. It will also be possible to link the electronic assessment with medical records. Our results could help to select the most performing questions according to mental disorders or patient profiles, which, in turn, could be used to create shorter and more efficient questionnaires. We can see in our study that the question about medication intake had very low informative value.

There are still many concerns regarding e-mental health that need to be addressed. One of the main concerns reported by both professionals and users is related to the privacy, ownership, and responsible use of medical information [38,39]. This is one of the major challenges that eHealth needs to address by means of privacy-preserving technologies [40]. Accessibility and difficulties to find reliable sources of medical information are also important concerns in the population, especially among older adults [41]. Medical professionals also have doubts about the capacity of online information to improve the knowledge of patients and have reported concerns regarding the capacity of telemedicine to enhance physician-patient bond [39]. All these concerns must be addressed in order to improve the acceptability and use of eHealth tools. A recent study suggested that there is still a low preference for the use of eHealth tools among the adult general population [17]. However, those who have already used eHealth apps usually feel confident to continue using them. Some studies have reported a sense of security and the existence of a relational bond between eHealth apps and patients with psychiatric diseases [42,43]. Our analyses show that machine learning can help to classify e-mental health users and provide clues for their diagnoses and, importantly, their needs in terms of treatment. If machine learning helps physicians to take clinical treatment decisions based on data, the social perception about available eHealth tools will certainly improve.