Can AI help in tackling mental health issues?

A psychologist and her patient in her office

The use of AI for clinical care has started to find its feet, with medical imaging and triage being some of the first areas where ordinary people have felt the impact. Various startups have also latched on to the idea of using AI to solve for bottlenecks in heathcare. The hype around these tools has started to become something more than hot air: we are starting to see tangible results.

The optimism is well-deserved: healthcare has, for lack of a better way of putting it, been stuck in the 20th century. Don’t get me wrong, the 20th century was a massive upgrade over the centuries that came before. As legendary as Galen was, I would prefer to be diagnosed by someone with more modern sensibilities and training. However, healthcare has been rate-limited by a very critical node in its chain. The doctor. While doctors have been helped by tools which boost their ability to more granularly understand a sick patient, very few tools have aimed to reduce their cognitive load. Improvements in information technology have led to the elimination of the typewriter, the introduction of the electronic spreadsheet, improvement in calculation, the smartphone, and many other conveniences our ancestors would have been astonished by.

But for doctors, information technology’s impact on their work has been to give them better eyes and ears. A sick person time-travelling from the 1950s would marvel at our diagnostics, but she would fundamentally be cared for in a manner not dissimilar to her own time. By a doctor at her desk.

Until now. The advent of AI does not threaten to do to doctors what the word processor did to the typist. The disruption staring doctors in the face is akin to dictation software vis-à-vis the keyboard. While some people prefer to dictate their thoughts, a large number of people will still type everything out the old fashioned way. Similarly, some doctors will still prefer to read X-rays on their own. But we may be headed for a future where computers help the majority of doctors read X-rays instead.

The role of AI in mental health, on the other hand, is more nuanced than a cursory glance might suggest. Diagnosis in the case of, say, heart disease is much more algorithmic in comparison to a diagnosis of autism. While the specific parameters may vary, heart disease tends to look quite similar in populations around the world. A doctor who has treated a Caucasian in New York would have little trouble in translating her knowledge to an Inuit patient.

However, mental health poses a whole different challenge. While some diagnoses are objective (the MRI of patients with advanced frontotemporal dementia should have similar tells, for instance), depression may be far harder to get right. A doctor who grew up and received her training in New Hampshire may find it hard to get a grip on the cultural mores of New Delhi, thereby missing signs of depression in her Indian patient.

AI has both natural advantages and disadvantages here. Its natural advantage shines through in places where large datasets make training and analysis easier. On the other hand, large, diverse datasets may lead to tricky situations.

Cognitive assessment seeks to assess the functioning of the brain in thinking, language, using judgement, and memory. It can be used to diagnose multiple conditions:

• Dementia and other kinds of old age-onset disorders are often diagnosed using cognitive assessments. Their progression is also monitored using cognitive tests.

• Brain injuries can often be diagnosed this way.

• Cognitive impairment caused by malnutrition (such as a B12 deficiency), thyroid issues, depression, etc.

• Learning disabilities and disorders can also be diagnosed and cognitive assessments may be used to create individualised learning plans tailored for each child.

There are many other diagnoses cognitive assessments may help us in. However, a typical problem associated with them is subjectivity in their interpretation. Two doctors treating the same patient may judge a patient’s cognition to be very different from each other, given an identical set of responses.

This may make sense given the very contextual nature of cognition. As the famous dictum goes, you shouldn’t gain the measure of a fish by its ability to climb trees. But the inaccuracy of these processes has been such that the introduction of AI to help in making more objective judgements is being seen as a positive development. From Thakkar et. al. (2024):

Screening of cognitive deficits or impairments and early intervention is currently the most widely accepted strategy to manage a number of psychological disorders. The diagnosis of these is established through thorough assessments, which may also help in understanding cognitive pathophysiology. However, lack of proper standardized screening and guidelines often leads to undiagnosed cognitive impairment which further leads to increased disease progression and cognitive decline.

Automating the assessment and prediction process is the key to timely diagnosis and management. The advent of AI has resulted in automated assessment techniques which improve the accuracy of diagnosis. ML and AI-based approaches like Support Vector Machine (SVM), neural networks and ensemble techniques like Convolutional Neural Network (CNN), AlexNet, GoogLeNet and LeNet5 have yielded some of the best results and accuracies when it comes to the use of AI for the assessment of cognitive mental health disorders.

The evolution and adoption of AI has the potential of leading to early and more accurate diagnosis of cognitive impairment and, perhaps, more effective diagnosis of the underlying cause such as Alzheimer’s Disease and the like. Several startups work in this area:

• Limbic: They offer AI tools like Limbic AccessAI, which claims to speed up assessments by 50%, saving time and improving accuracy for mental health providers. It is used by the NHS, UK.

• EllipsisHealth: This startup uses voice-based AI, called Sage, to check mental health by analyzing speech, helping identify issues like anxiety.

• Kintsugi: They also use voice AI for daily mental health check-ins, helping diagnose conditions like sadness, with over 200,000 users worldwide.

These include disorders like Down’s Syndrome, AHDH, Autism, etc. In addition to cognitive assessments, AI can be used to analyse patient data to accurately diagnose these disorders. The examination of neuroimaging data, for example, or eye-movement patterns can be used to gauge a number of disorders. Based on this, AI systems have been developed to help physicians, or even ordinary people, in the diagnosis of multiple development disorders.

None of these tools have yet seen widespread adoption. However, that may change with time as AI systems learn and adapt and as news of these products spreads. The primary reason, again, as with cognitive assessments, is to provide more objective and accurate measures of conventional intellect and development.

While AI shows immense promise in this field, it's important to remember that these AI tools are primarily designed to be screening and assistive tools for clinicians, parents, and educators. They are intended to aid in early identification, provide objective data, and streamline workflows, but they do not replace a comprehensive clinical diagnosis by qualified medical and developmental specialists. The ultimate diagnosis still relies on expert human evaluation.

Neurodegenerative disorders are characterised by a gradual reduction in the number of neurons in the brain. The brain may lose neurons from different regions, giving rise to different diseases such as Alzheimer’s Disease, Parkinson’s Disease, ALS, etc.

The oldest form of AI development in this area has been a more traditional and objective approach: interpreting brain scans and EEG data. Various biomarkers can also be identified with AI, which, when combined with other data, can lead to accurate and early diagnosis of various conditions. This includes analyzing cerebrospinal fluid for proteins like tau or beta-amyloid in Alzheimer's, or genetic markers associated with Parkinson's.

However, the holy grail of AI intervention in this area lies in a holistic interpretation of patient biomedical data. AI models analyze historical patient data, genetic information, lifestyle factors, and clinical records to predict disease progression and identify individuals at high risk. This enables proactive interventions, such as starting treatments earlier or planning care strategies. Predictive analytics can estimate how quickly a patient's condition might deteriorate, allowing for timely adjustments.

The intersection of AI and humans and their mental conditions has led to the development of emotional AI, where technology has been designed to learn, perceive from and respond to human emotions. The field is quite nascent right now, but has great potential going forward. The technology has many potential uses both in the personal and professional spheres. Emotional sensing, in particular, is an area with many potential benefits, if executed properly. Again, from Thakkar et. al.:

Emotion sensing, a pivotal aspect of Emotional AI, traces its origins to affective computing in the 1990s. Enabled by weak, narrow, and task-based AI, Emotional AI aims to comprehend and interact with emotional states by analyzing a spectrum of data related to words, images, facial expressions, gaze direction, gestures, voices, and physiological signals, such as heart rate, body temperature, respiration, and skin conductivity. The input features for emotion recognition could include facial expressions, voice samples, or biofeedback data, while the output encompasses emotional states used for various purposes. Common machine learning techniques like convolutional neural networks, region proposal networks, and recurrent neural networks are frequently employed for these tasks.

While this may sound like the unlikely beginnings of a 1930s dystopian novel, the technology itself is progressing quite fast and is likely to see mass deployment soon. The advantages of such an approach are many. While areas such as cognitive assessment/therapy and neurodegeneration are seen as medical conditions to be treated, the application of AI techniques for helping out in general mental health may have the potential to be even more transformative. Since the biggest problem with mental health issues is lack of access to care or a wish to avoid social stigma, being able to give people means to access care without being judged might be the push they need to take their mental health seriously.

AI may also be able to help individuals understand their emotional states better. While a certain degree of cultural and situational understanding is required to accurately gauge an individual’s emotional state at any given time, the more rigorously methodological and empirical way in which AI tends to both gather and analyse this data has the potential to give a user a better and more granular understanding of their mental state at any time. Tying this analysis to a service, for example, a guided meditation class based on the data gathered about your emotional state at the time, may help people lead more emotionally fulfilling lives, especially those dealing with emotional dysregulation and/or mood disorders.

Emotional AI may end up with more benefits than merely(!) general mental health. Creative endeavours are very emotion-forward tasks, but ask any artist about inspiration, or the lack thereof, and you will get a cavalcade of thoughts about their recent dry spells. This has inherently been seen as part of the creative process, but artificial intelligence is inherently a different form of intelligence than any other we have seen before. It may have ways to break through creative barriers which ordinary humans may not. The current means of interacting with AI is through chatbots. As AI becomes more integrated with healthcare devices which are always on us (such as smartwatches), the data being collected from these sources may have the potential to inform interventions in this area as well.

However, this area, more than any other, requires deeper thought and engineering. Emotional states are complex topics, and AI-algorithmic biases are more difficult to detect and solve for than many other programming issues. As Thakkar et. al. put it:

Mental health disorders are complex and often involve a combination of subjective symptoms, environmental factors, and personal histories. AI algorithms may struggle to accurately diagnose conditions that require nuanced interpretation of these factors. The human aspects and reading in between the lines of a patient's history may be missed by algorithms. Mental health disorders can manifest differently in different individuals. This variability challenges the development of a one-size-fits-all AI solution and requires algorithms that can adapt to diverse presentations. AI algorithms can produce false positives (diagnosing a disorder that isn't present) and false negatives (failing to diagnose a disorder that is present). AI algorithms may lack the ability to fully understand the context and nuances of a patient's life, emotions, and experiences, which can affect accurate diagnosis and assessment. Many mental health disorders have overlapping symptoms, making accurate diagnosis challenging even for experienced clinicians. AI algorithms may struggle with this complexity as well.

While these are not impossible problems to solve, they require time, effort, and patience. Mental health problems have been medicine’s bugbears for a while now. Hopefully, new technology can lead to new treatments and new gains.