Workshop Summary & Outcomes
The workshop formed part of the ACII 2019 conference in Cambridge (UK), and was well attended.
It had an extensive program that started with a keynote on digital mental health services in the US; followed by multiple presentations of accepted workshop papers;
invited talks of established researchers in the field; group discussions; and conversations about future steps to continue experience sharing within, and growth of,
the wider research community on ML4AD.
Talks & Activities
Topics and applications areas ranged from detections of depression from body movements (Daoudi et al. 2019) and online suicide risk predictions on Reddit (Jones et al. 2019),
to various approaches to assist stress recognition (Giannakakis et al. 2019; Nakashima et al. 2019; Taufeeq Uddin & Canavan 2019), e.g., through biofeedback such as HRV (Giannakakis et al. 2019);
as well as one study of an impulse suppression task to help detect people suffering from ADHD (Leontyev et al. 2019);
and strategies for generating better ‘wellbeing features’ for end-to-end predictions of future wellbeing (Li et al. 2019).
Discussions of the workshop papers, keynote and invited speaker presentations foregrounded many important challenges in the field.
Amongst others, they related to: (i) international difference (e.g., in how mental health services are compensated; cross-cultural values);
(ii) the importance of supporting the interpretability of ML systems; (iii) defining, selecting or developing good quality mental health measures;
(iv) ethical issues when developing or deploying ML applications; and also challenges in (v) data processing and feature extraction.
Discussion Groups
Towards the end of the workshop, attendees formed discussion groups to dive deeper into conversations about topics of ‘measurement’ and ‘ethics’.
Topic: Measurement Challenges
The topic of measurement and associated challenges frequently occurred throughout the day and ranged from described difficulties in labelling data and establishing
‘ground truth’; to limitations involved in defining, selecting or developing specific mental health measures and targets;
and also related ethical considerations as to what may be considered as ‘safe’ measures to administer to study participants,
or people who are perhaps self-managing a condition as part of everyday life. Two additional areas of debate received particular attention:
1. How to conceptualize mental health states/ symptoms (e.g. categorial vs. dimensional models)?
In terms of research foci and the types of ML tools that we develop, there appears to be a trend on detection and diagnosis,
which may partly be explained by the availability of data and measurement tools in this space. This informs how research targets are shaped.
At the moment, the trend seems to be to try and match the data about a person to a diagnosis (e.g. broad category of depression) –
whereas the more important task or goals might be the mapping of the person to a relevant treatment. Furthermore, if we look at mental illness such as depression as one broad category,
we may not be able to account for the many variations of depression combinations and factors; meaning that building a model for monitoring depression is only vaguely defined.
This suggests perhaps targeting simpler definitions of the problem(s) that we are trying to solve.
Related to these discussion points were key questions such as:
What is the health/ medical problem that we are trying to address? Are we asking the right questions?
How to ensure that the (often complex) models/ solutions we develop in computing science really meet a clinical need? What are ‘the right’ use cases for ML?
How to define/ select/ develop good quality measures?
2. The value of objective vs. subjective assessments – do they need to compete with each other?
Excitement about often passively and continuously data captured about peoples behaviours through multimodal sensors, including biofeedback,
or content they create online started to shape perceptions of new, advanced data approaches to provide ‘more objective’ insights;
especially when compared to other more subjective methods such as self-report questionnaires. Discussions pointed out that we cannot strictly define what is subjective or objective.
Thus, instead of looking at these approaches in competition, perhaps a more promising route is to look at interesting relations that come about through the combination and triangulation of
different methods; and what each may say about the person (e.g., questionnaires about stress say one thing; the physiological data says something else – how can this be explained).
Related to this were conversations about the role of clinicians and how ML systems can empower them?
For example, while we can detect schizophrenia from mobile phone data and clinicians would evaluate such data as potentially useful,
it can be difficult for clinicians to make sense of a lot of data that patients may only share when presenting to them once a month.
For these kinds of scenarios it is really important that the data in presented in ways that is helpful to clinicians and can complement them in their work practices,
rather than being overwhelming. Related to this were questions such as:
How can we empower clinicians through data tools (what is helpful, relevant, actionable)?
How can we help clinicians to trust us?
How can the results of ML help make concrete actions/ interventions for clinicians/ patients?
Topic: Ethical Challenges
Inevitably, when discussing the role of ML and possibilities of ML-enabled interventions for use as part of real-world mental health services,
the conversation would turn to ethical issues. The following two themes were most prevalent:
1. (How) Should we (or not) communicate ML-detected/ diagnosed mental health disorders, or risks?
A key conversation topic evolved around communications of ML-detected or diagnosed mental health disorders or mental health risks
(e.g. of suicide on social media) to people, who may not have been aware of a mental health problem, or risk.
On the one hand, being able to detect problems (early) can help raise awareness, validate the person’s experience,
invite help-seeking and a better management of a condition. On the other hand, people may not necessarily be aware of their data traces
being analysed for the purposes of mental health screening, or may not want to be ‘diagnosed’ with a psychiatric condition due to associated stigma,
and also potential negative effects on their personal or work life. For example, a diagnosis of a mental disorder can have severe implications
for people in work professions such as the police force, or firefighting. Thus, how to balance both peoples “right to be left alone” and “right to be helped”?
Related to these discussion points were key questions such as:
How to sensibly communicate the detection/ diagnosis of a mental health problem or disorder?
Should passive data only be collected and used for self-reflection and self-care of the person?
How to show risk factors to people in ways that is actionable (e.g., a diagnosis alone may not be helpful unless the person knows what they can do about it)?
What kinds of behavioural interventions should perhaps not be developed or tested with people in-the-wild (e.g., without appropriate safeguards in place)?
2. What are the broader implications of ML-interventions and how to reduce risks of misuse?
It is hard to predict what unanticipated consequences a new ML-enabled mental health intervention might have on the person, their live, or society at large.
Partly this is due to the way in which we tend to study well-defined problems whose solutions may not transfer to other contexts outside of those for which they’ve been designed,
or trained for. For example, in the context of developing an emotion recognizer that was based on a person’s facial expressions,
it was discussed what the implications might be, if someone was re-purposing this technology e.g. to identify children who are not working enough at school,
or employees who appear to be less productive at work.
In terms of misuse, examples discussed further included: difficulties to prevent the (mis)use of developed tools with low clinical accuracy in clinical practice;
and challenges related to user consent and data control, where it may not be possible or transparent for users to know who has (and owns) information about them;
how data is used, re-used, or how they could take their data back? Thus, key questions included:
How to responsibly design and develop ML systems?
How can we help reduce the risk of misuse for the technologies we develop?
How to re-think consent processes and support user control over their data?
Responding to the question how we can – in practice – reduce risks of our tools being misused, or having negative impact, Rafael Calvo also pointed to a psychological framework
that suggests four sensitives, and how they can be addressed:
Anticipatory: interdisciplinary work
Reflective: slow down and think
Deliberative: participatory work
Responsive: notice and adapt
