Algorithms and machine learning (ML) have conquered, transformed, and essentially revolutionized people’s everyday lives in many aspects (eg, with personalized Google searches, self-driving cars, or convenient smartphone apps) [1,2]. In recent years, self-diagnosis apps have emerged that allow patients to look for a diagnosis based on entered symptoms [3,4]. ML is already being standardly used in various applications like estimating diagnoses from radiology images [5], but the adoption and acceptance of new technologies in health care in general is curbed by trust issues, strict regulations, and lack of thorough investigation [3,6]. Little testing of the aforementioned apps has been previously performed; Semigran et al tested self-diagnosis apps in general in 2015 but did not mention ML as an underlying technology [7]. A recent scoping review by Aboueid et al in 2019 named several apps of this type [3], but only two have been tested in their diagnostic functionality so far [8-10]. The US Food and Drug Administration (FDA) had previously excluded “symptom checker” apps from the enforcement of the strict rules that are usually applied to medical devices [11] but has lately released a white paper with a proposal for possible changes in the regulation of self-diagnosis apps, introducing a new discrimination between “locked” and artificial intelligence (AI)–based/ML-based learning algorithms, with the latter falling under a stricter set of rules [12]. However, the provided definitions of the two categories are still vague and rely on the manufacturer admitting the use of AI-based technologies rather than on hard criteria. Repeated examinations of such apps would offer insight into whether the results they offer are stable over time or change, which would also indicate actual use of “learning” AI technologies by the manufacturers, rendering the apps more than plain symptom checkers. Thus, the first aim of this study was to test apps that have previously been mentioned in the scientific literature on a set of diagnoses and to subsequently follow up on the results after a deliberate time interval to determine whether the algorithms change over time. A significant change in the results might indicate that learning algorithms are used by the manufacturers.

None of the aforementioned studies had tested self-diagnosis apps in the field of ophthalmology, the author’s primary specialty with over 12 years of clinical experience. Thus, the second aim of the study was to test the efficacy of self-diagnosis apps by presenting them with three common diagnoses representing three levels of urgency in this field to see if their results and treatment recommendations match those of someone familiar with the topic.

During testing of the apps, some notable observations could be made about their behaviors. Ada seemed to ask redundant questions in the end, (eg, asking about eye pain when this was the primary symptom entered at the beginning). It could be speculated that this functionality serves to add input for the diagnosis to the database. Ada circumvents the common “black box” problem in ML [32] by offering a pictorial description of how many patients in 10 with the given symptoms have the suggested diagnosis. This additional information seems to indicate that by Ada’s accounts, the provided symptoms for dry eyes seemed to correspond less with the diagnosis in 2020 (“8 in 10” vs “5 in 10 people”). Here, one could speculate that there are problems incorporating the data the app accumulates over the years. Generally, the provided statistics indicate that either Bayesian probabilities are used in some way, as artificial neural network (ANN) output activities are not linked to statistical values, or the offered values are interpolated from ANN outputs. The former assumption seems to be backed up by information published by Ada Health itself, where the use of Bayesian networks is mentioned [33]. Buoy asked several questions that seemed off topic (eg, for the user’s health insurance), and in 2018 it presented users with pictures of medical conditions for comparison that may not be suitable for laymen, like testing a patellar reflex in the dry eye case or comparing one’s cornea with a microscopy picture to identify Horner-Trantas dots. In 2020, no pictures were offered for the same set of symptoms. In both years, both Babylon and Buoy failed to produce a useful diagnosis and also gave out very few treatment recommendations, with some results being very far off, such as “Bone issue” or “Non-bacterial brain inflammation” diagnoses by Buoy for the retinal tear patient, contributing to the overall result that Ada and Your.MD fared significantly better in the test than the other two. The variety of treatment recommendations given by the apps for the same starting sets of symptoms is also remarkable. Ada made it simple by generally recommending emergency care for virtually every diagnosis, which may help the manufacturers shift responsibility to the patient but counteracts the possible value of good medical advice. Babylon seemed to send any patient who chooses “severe pain” as a symptom to the emergency department, which is a good outcome for the glaucoma patient, but no diagnosis was given, and its other recommendations were very general. In 2020, Buoy gave the glaucoma patient the advice to seek medical advice within three days as its first option, followed up by “emergency treatment” as second and third options, which would confuse a real patient. Your.MD provided the most valid recommendations in this study, but also worsened on the dry eyes diagnosis from 2018 to 2020, now unnecessarily transitioning from self-treatment to emergency care.

While the number of questions the apps asked did not significantly change between the years, the temporal variances in diagnoses and treatment recommendations indicate the use of learning algorithms in all four, suggesting that the algorithms used for history-taking and diagnosis calculation are changing over time and would thus fall under the FDA’s proposed regulations for learning or “nonlocked” algorithms. In terms of their effectiveness in diagnosing ophthalmic diseases, the results were mixed with a tendency to worsening. It is noteworthy that no trend to improvement of history-taking and results could be observed at all. On the contrary, Ada and Your.MD worsened in their diagnostic outputs, while Babylon and Buoy were stable at a low level. This deterioration of diagnostic performance seems to contradict the very purpose of using “learning” algorithms in the first place and certainly justifies further inquiry. It is also notable that while two of the apps ask more questions than the other two, there seems to be no correlation between the number of questions asked and the quality of the results. On the contrary, the app with the highest overall score had the second-lowest total of questions asked. This indicates high variation in their diagnostic approaches and efficiencies, all worthy of subsequent systematic evaluation. Their algorithms are undisclosed; judging by the apps’ workflows, they all basically resemble the adaptive feedforward neural network–based mobile diagnosis engine that the author conceptualized in 2016 [34], which in its framework resembles the classic AI game “20 Questions” [35]. In both frameworks, two separate neural networks (or similar algorithms) separately calculate the current diagnosis based on the available information and the next best question based on the input up to that point. Where these examples used simple ANNs, the apps may also utilize the previously mentioned RNNs, Bayesian networks, or convolutional neural networks [15], accessed through the chat bot.

There are several limitations to this study. First, the cases that were entered included information that seemed irrelevant to the author for deciding on a suspicion of a potentially dangerous diagnosis; this includes, for example, the prevalence of diabetes, a history of smoking, and seemingly unrelated questions. It is possible, however, that to a large database that collects and sorts information without bias, answering the questions from a biased physician’s point of view might in fact mislead the algorithms. Second, the evaluation of the results is as subjective as entering the symptoms, which ultimately might test the apps’ ability to imitate a potentially flawed physician rather than whether they can correctly identify diagnoses. This could be improved by introducing systematic evaluations in the framework of a randomized controlled trial. Semigran et al had used human input and output on a randomized stack of diagnoses to assess self-diagnosis apps [7]. New methods may be necessary to investigate AI-driven apps in the future, possibly including some degree of automation considering the superhuman data storage capacities such systems can house and taking into consideration the dynamic of the algorithms. A simpler but also less systematic approach would be to include more physicians in the evaluation of the apps and average their assessments as has been done before [10]. The possibility of the manufacturers adapting to known sets of questions (eg, from this study) should also be considered upon further investigation. Third, the sample size is low. In future investigations, large-sample investigations should be preferred. Other authors like Fraser et al in 2018 have already demanded standardized and transparent procedures for examining such devices [36]. In 2019, Kelly et al advocated for a focus on peer-reviewed studies in order to increase trust in AI devices and added that the introduction of consumer-oriented technology offers the opportunity for vast prospective studies with the new collected data, provided that a sufficient level of data transparency is reached [37]. They also mentioned that an extension to the existing TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) statement, which has defined recommendations for the evaluation of diagnosis prediction since 2015 [38], is being developed to include ML algorithms [39]. In addition, the World Health Organization and the International Telecommunication Union are working on benchmarking frameworks for AI tools in health care [40]. All of these could act as guidelines for future scientific exploration of the topic, but they will require funding and manpower. The four enterprises mentioned in this paper are employing physicians, mostly with an additional formal or informal education in medical informatics or similar, at high ranks within their hierarchy, in two of them even as cofounders [41-44]. Now that this Pandora’s box of AI in the hands of patients and corporations has been opened, the question is whether this novel type of physician that supervises and administrates an automated diagnostic system will be mirrored by scientific counterparts who publicly evaluate the apps’ performances, or whether these essential data will remain undisclosed—a common practice in the commercial sector due to conflicts of interest. Considering the apps’ possible leverage and impact on public health [4], this should be in the public interest. The prospect of AI support for physicians provided by simple and accessible apps in the hands of layman users could be a golden one, as long as they actually learn and improve. Most importantly, they need to satisfy the crucial premises within the field of health care: to be efficient and safe.