https://orcid.org/0000-0002-3779-5076
![Sample our Health and Social Care journals, sign in here to start your FREE access for 14 days]({"type":"image" , "src" : "/sda/5938/TOC-Subject-Sample-2021-Health-Social-Care.jpg"})
Abstract
Well before COVID-19, there was growing excitement about the potential of various digital technologies such as tele-health, smartphone apps, or AI chatbots to revolutionize mental healthcare. As the SARS-CoV-2 virus spread across the globe, clinicians warned of the mental illness epidemic within the coronavirus pandemic. Now, funding for digital mental health technologies is surging and many researchers are calling for widespread adoption to address the mental health sequelae of COVID-19. Reckoning with the ethical implications of these technologies is urgent because decisions made today will shape the future of mental health research and care for the foreseeable future. We contend that the most pressing ethical issues concern (1) the extent to which these technologies demonstrably improve mental health outcomes and (2) the likelihood that wide-scale adoption will exacerbate the existing health inequalities laid bare by the pandemic. We argue that the evidence for efficacy is weak and that the likelihood of increasing inequalities is high. First, we review recent trends in digital mental health. Next, we turn to the clinical literature to show that many technologies proposed as a response to COVID-19 are unlikely to improve outcomes. Then, we argue that even evidence-based technologies run the risk of increasing health disparities. We conclude by suggesting that policymakers should not allocate limited resources to the development of many digital mental health tools and should focus instead on evidence-based solutions to address mental health inequalities.
ACKNOWLEDGMENT
Thanks to Walter Sinnott-Armstrong, Jana Schaich Borg, and members of MAD Lab for their helpful feedback on earlier drafts of this manuscript. Thanks also to two anonymous referees. Their careful comments helped to strengthen and clarify our arguments.
DISCLOSURE STATEMENT
No potential conflict of interest was reported by the author(s).
Notes
1 In their (2019) guidelines on digital interventions, the WHO defines digital health as “a broad umbrella term encompassing eHealth [electronic health: ‘the use of information and communications technology in support of health and health-related fields’], which includes mHealth [mobile health: ‘the use of mobile wireless technologies for health’], as well as emerging areas, such as the use of advanced computing sciences in ‘big data’, genomics and artificial intelligence.” Given the broad scope, this terminology best captures the variety of approaches in digital mental health as well.
2 We should also note at the outset that developments in this field proceed at a breakneck pace, much faster than the traditional academic publishing cycle. In our experience researching this topic, it is not unusual that a digital mental health application will be introduced on the market, and then be re-branded, bought-out by a competitor, or disappear altogether within the timespan of a few months. Thus, our examples in this section are best understood as a snapshot of DMH technology from a particular point in time (namely, Summer/Fall 2020). Nonetheless, we made every effort to ensure our analyses will remain relevant to future developments in this area.
3 Of course, one of the defining features of the COVID-19 pandemic has been the rapid shift and adoption of tele-health generally, and tele-mental health specifically (e.g. use of phone, text, or video communication involved in the delivery of mental health services on platforms such as Doxy, Teladoc, or Mend). Before COVID-19, these services were often promoted as an effective means of providing mental health services to low-resources areas (Kaonga and Morgan Citation2019) and traditionally under-served populations, including rural areas (e.g. Myers Citation2019; Speyer et al. Citation2018), indigenous communities (Hensel et al. Citation2019), and geriatric patients (Gentry, Lapid, and Rummans Citation2019). While tele-mental health certainly fits under the broad umbrella of DMH, in contrast to some smartphone apps, many standard tele-health approaches lack the scalability required to address the increasing demand for mental health services (see Section 3.3 below for further discussion).
4 For example, according to SensorTower the world’s top 10 combined English-language mental wellness apps “accumulated close to 10 million downloads, up 24.2 percent from the installs they generated in January 2020” (Chapple, Citation2020). It of course remains to be seen if these trends will continue after the pandemic, but this rapid increase is undoubtedly significant.
5 Digital phenotyping is described as a family of “approaches in which personal data gathered from mobile devices and sensors are analyzed to provide health information.” (Martinez-Martin et al. Citation2018, 1). According to Insel, smartphones provide “an objective, passive, ubiquitous device to capture behavioral and cognitive information continuously,” with the potential to “transmit actionable information to the patient and the clinician, improving the precision of diagnosis and enabling measurement based care at scale” (Insel Citation2017, 1215). In turn, these approaches promise to “revolutionize how we measure cognition, mood, and behavior,” and “transform the diagnosis and treatment of mental illness globally by enabling passive, continuous, quantitative, and ecological measurement-based care” (Martinez-Martin et al. Citation2018, 4).
6 Interestingly, the game was in regulatory limbo for the past two years waiting on a decision from the FDA. But by the time the green light was given on June 15, 2020, the game was already available online, due to the FDA’s decision in April 2020 to relax regulations on low-risk mental health devices during the COVID-19 pandemic.
7 For example, Medicare quickly modified its policies to allow clinicians to use non-standard telehealth technologies such as FaceTime or Skype (Wilser Citation2020) and also allowed clinicians to bill for them (Figueroa and Aguilera Citation2020). Likewise, the Health Insurance Portability and Accountability Act (HIPAA) Rules, immediately expanded the remote communication channels that health care professionals can use with patients, even if such channels do not fully comply with HIPAA (HIPAA Office for Civil Rights (OCR) Citation2020). Many states also waived the requirement of requiring psychiatrists to provide services only to patients within the states they are licensed to practice in, so they can now provide such services out of state (Gautam et al. Citation2020).
8 For example, a recent meta-review by Lecomte et al. (Citation2020) reported retrieving over 2,500 potential papers and 24 meta-analyses related to DMH.
9 Following Cohen (Citation1992), effect sizes around 0.8 are considered large, effect sizes around 0.5 are moderate, and effect sizes around 0.2 are small. There are, however, perennial debates about the relationship between such measures of statistical significance on the one hand, and clinical significance on the other. In terms of interpreting the meta-analytic results reported here, a concrete example may be helpful. The 17-item Hamilton Depression Rating Scale, with a range from 0 to 52 points, is “the most commonly used depression rating scale and is the recommended scale by psychiatrists worldwide” (Jakobsen, Gluud, and Kirsch 2020, 2). A decrease of three points on this scale (e.g. a score of 47 at baseline, and then a score of 44 after an intervention) corresponds to a standardized mean difference of 0.5. A drop of seven points on the scale corresponds to an effect size of around 0.8. For more context, Hieronymus et al. (Citation2020) estimate that commonly prescribed antidepressants have an effect size of approximately 0.3 compared with placebos.
10 After all, if one signs up for a research study about the effects of a smartphone app on mental health, but one never uses an app (as in a waitlist control), or one just listens to music (as in an “active” control), it would not be difficult to determine that one was not in an intervention condition.
11 For a striking example of this gap between machine prediction and clinical intervention, see Elish & Watkins’s (Citation2020) in-depth study of the Sepsis Watch AI tool deployed at the Duke Health. Their analysis shows that even with (relatively) straightforward conditions like sepsis, there exist tremendous difficulties in translating AI-driven predictions into improved clinical outcomes. Such difficulties are likely to be even more pronounced in the mental health context.
12 One might object here that such NLP projects are more aimed at research than clinical application. However, when NLP researchers claim, as they often do, that such research can, for example, “greatly facilitate targeted early intervention” and “provide previously unavailable information for clinicians on which to base treatment and prognostic decisions” (Bedi et al., Citation2015, 6), it seems fair to raise this criticism, especially when the myriad difficulties of translating computational research into improved psychiatric outcomes are not discussed in any depth. See Velupillai et al. (Citation2018) for an analysis of the many difficulties (and also opportunities) of translating between NLP and clinical outcome contexts.
13 Here, they reference Yellowlees and Shore (Citation2018). Studies examining the efficacy of telehealth via phone- or video-calling vs. face-to-face therapy tend to point in the same direction for depression (Berryhill et al. Citation2019), PTSD (Acierno et al. Citation2017), and the use of internet-based cognitive behavioral therapy more generally (Andersson et al. Citation2019).
14 In fact, these forms of tele-health may even be less time- and cost-effective than traditional face-to-face therapies, at least at first, given various struggles with internet connection issues, scheduling, interface functionality, and the like. Thanks to an anonymous reviewer for suggesting this point.
15 To reiterate a point made above, evidence-based considerations do support the limited use of smartphone apps in stepped-up care settings, for example, and the same point might be made about the use of these technologies as “stopgap” solutions amid a public health emergency. This is fine as far as it goes. But we worry that many DMH proponents are advocating for much more than these short-term, emergency measures. Anticipating the themes of the next section, there is a real risk that, without persistent critical oversight, what starts as a short-term, “stopgap” solution slowly becomes the “new normal” in such a way that existing health inequalities are exacerbated: The relatively well-off in society get evidence-based, face-to-face therapies while the less well-off get automated chatbot therapists. Indeed, in a recent op-ed, Green (Citation2020) observes precisely this dynamic, albeit in the context of primary education, in proposals for so-called “learning pods.”
16 These worries are pronounced in other forms of DMH as well, especially NLP applications which are trained only or primarily on English speakers’ voice, text, social media posts, etc.