AI in Public Mental Healthcare: Madness or Genius?

AI in Public Mental Healthcare: Madness or Genius?

People tend to resist technology they find intimidating. This could presumably be a barrier for AI, especially in fields where it is difficult to imagine how machines and humans could outperform humans or machines on their own. However, if applied well, AI holds the promise of radical change. For instance, being mindful about climate change, according to MIT Professor Daniela Rus, employing an already existing AI algorithm to direct taxis around New York could reduce the number of necessary vehicles from 14000 to 3000.

A new AI-based approach was tasked with reviewing images of lymph node cells to diagnose cancer. On its own, it had an error rate of 7.5%, worse than the 3.5% rate of human pathologists. But when both the AI system and the pathologist reviewed the data, the error rate went down to only 0.5%. – Daniela Rus, Professor at MIT

Similarly, many companies are working on AI enhancements for commonly used equipment that have remained unchanged for decades. One example is weed-killing robots with computer vision systems that can distinguish weeds from crops as they roll through fields. The machines reportedly use 20 times less herbicide than standard methods that typically involve blanketing entire fields with chemicals linked to negative health outcomes. The logical extension of such an innovation would be to integrate it into machinery farmers already use, such as tractors. This article investigates whether machine learning, or deep learning, could be employd in mental health care – an industry where the successful combination of humans and machines would be an absolute imperative for implementation.

Artificial intelligence (AI) is currently unused within mental health services. This should hardly come as a surprise to anyone, as it is particularly difficult to imagine how machines could be useful within the field of psychiatry/psychotherapy. To provide value, a “cold algorithm” would need to facilitate fruitful human interactions – a feat almost impossible to imagine at face value. However, our mental imagery could perhaps be informed by one study demonstrating that the combination of humans and machines resulted in a significantly improved diagnostic accuracy within somatic medicine. In this study, human radiologists performed diagnostics with an error rate of 3.5%, while machines had an error rate of 7.5%. Interestingly, the combination of humans and AI resulted in only 0.5% error. This finding allows us to wonder whether the combination of people and computers could outperform humans on their own. If so, then the implementation of AI could prove beneficial when it comes to the diagnosis (and/or screening) of psychiatric patients.

This has the potential to improve mental health care significantly – both by making treatment more efficient and more precise, i.e., tailored to the specific patient's needs.

Before we allow our inner artist to sketch out such a desired future, we need to investigate the necessary ingredients, or preconditions, for such a reality to actually manifest. First, in order for an ML algorithm to reach sufficient predictive precision, we would need enough diagnostic data for it to learn from. As there are diagnostic data of hundreds of thousands of psychiatric patients in secure research data bases around the globe, it seems reasonable to expect that there is enough data for an ML algorithm to be able to achieve adequate predictive accuracy. Next we should consider the nature of the phenomena we want the algorithm to learn about. Many psychiatric disorders are defined as significant deviations from cultural norms or expectations. Hence, this is a field where one size fits few. Consequently, the data would also need to be diverse, such that the algorithm would be exposed to cultural variations. As the human bandwidth is limited, an algorithm evolved from such data, would capture cultural nuances far beyond the capacity of any single human. If successful, the cultural aspect of the algorithm could therefore help clinicians when facing patients from unfamiliar cultures. Another concern is that diagnostic categories within mental health care are typically overlapping, and often exist as dimensions in relation to each other, rather than as discrete and distinct concepts. However, this should not prevent the algorithm from being able to learn to predict diagnostic categories: ML has been successfully utilized in several complex areas with overlaps and dimensional categories, such as in Google translate. Hence, if implemented successfully, such an algorithm could assist clinicians in diagnosing psychiatric patients with increased precision. Better diagnostic accuracy has the potential to make mental health services more specific in their treatment of different diagnoses, and also ensure that patients are provided treatments indicated by science. This has the potential to improve mental health care significantly – both by making treatment more efficient and more precise, i.e., tailored to the specific patient's needs.

Management and leadership scholars have spent the majority of the 20th century developing the principles of “people managing people”. The central management question of the future will be “How can we get humans and machines to best work together?” It is not about AI vs. people, it is about AI and people.

Except from both therapists and patients asking Google about symptoms and diagnosis, there has been a severe lack of implementation of (AI) technology within the field of mental health care. Hence, AI is currently unemployed in most mental services. This lack of technological adaptation also implies that there has been little increase in efficiency within this field over the last decades (with exceptions due to improved treatment skills). This stagnation has several causes, such as ethical concerns and the demand for privacy; disallowing personal data collection. There are numerous examples of applications that could be useful within this industry, such as providing patients with an app that would allow them to evaluate, and provide feedback to their therapists, or developing machine learning algorithms that could assist clinicians in diagnosing patients. Most organizations within this field rely heavily on cost leadership, while few actors have traditionally focused on creating uniquely desirable services. However, recently there are a growing number of cases were organizations focus on, and offer highly specialized services to a specific psychiatric population, e.g., eating disorders, obsessive compulsive disorder, substance abuse or personality disorders. In such cases the demand for a correct diagnosis is of great importance, as specialized treatments are tailored for, and target, different clusters of maladaptive psychodynamic patterns.

A correct diagnosis is the first step in any efficient treatment (for any disorder), and there are currently research initiatives within the EU to investigate how psychiatric diagnosis best can be standardized across geographical, demographical and cultural differences. Let us for example take a closer look at one diagnostic category associated with substantial societal cost: personality disorders (PDs). There are currently two competing paradigms for diagnosing PDs worldwide. Although both paradigms have their merits, they are currently also associated with important limitations. To be more specific, the diagnostic instruments accompanying these two models are lengthy and require a substantial amount of training. These issues may hamper diagnostic procedures in clinical practice, and have an adverse impact on the reliability and validity of ensuing scores. Current research aspires to develop a computerized adaptive test (CAT) for PDs, as an alternative to the diagnostic instruments used today. In such an adaptive test, the questions selected for inclusion in the interview are tailored to the responses of the individual patient, thus it is able to provide a more personalized assessment. However, such tests would not necessarily increase the precision of the diagnosis, as there is no feedback or input from the clinician, and the test is based on a fixed set of rules extracted from a research sample. It therefore seems reasonable to suggest that machine learning (AI) could better be employed in the task of assisting therapists in diagnosing PDs and other mental disorders.

As a first step to implement AI in mental healthcare at a larger scale, it seems advisable to perform a pilot project. Such a pilot should ideally involve a severe psychiatric disorder with (a) well-defined and documented treatment(s). It should also involve a psychiatric disorder, which is typically problematic to diagnose. We should not employ AI in order to identify substance abuse, and even though eating disorders certainly can be hard to identify at occasions, it often boils down to a pretty concrete calculation of body mass index. The treatment of these psychiatric disorders are complex, but it would be hard to argue well for spending considerable resources in order to ask AI for help in diagnosing these patients. Two categories come to mind when searching for the optimal candidate: Bipolar disorder and personality disorders. In both cases misdiagnosis can be fatal, and even experienced clinicians often have a hard time with these severe disorders. As a treatment for Bipolar disorder is typically medication (e.g., Lithium) we should choose personality disorders, as this disorder needs long-term psychotherapy, and hence the correct diagnosis of this disorder will be of great cost value. PDs can be treated with evidence-based treatments. Autism spectrum disorders (e.g., Asperger) can look very similar to PDs, but we have no well-documented treatment for this disorder. Hence, differentiating these two categories will be one challenge for AI. Even though public mental healthcare is not infused with competition, being able to develop an AI algorithm able to successfully diagnose PDs would be of great benefit for the organization involved. This would allow for further differentiation, and a potentially more secure and efficient treatment. We propose to implement a comprehensive research project, where all the available data for patients with PDs are given to a deep learning algorithm (statistical approach). At some point, this algorithm will have enough data to be able to divide a high dimensional space constructed by all feature vectors in two – hence making a prediction about the presence of a diagnosis or not. The algorithm should be fed all diagnostic instruments, and ideally also data from the journal and possibly also patients' description of their own situation. AI could automate the task of translating such documents into vectors. As soon as good enough predictive precision is established from historical records, the algorithm would be enabled to guide clinicians in their diagnosis, both (hopefully) benefitting the diagnostic accuracy, while at the same time letting the algorithm continue it’s life-long (process of) learning.

AI = machine learning (ML), natural language processing (NLP), and/or robotics. Deep learning is one form of machine learning.

The correct diagnosing of PDs, which is known to respond particularly well to specific treatment methods, could increase both quality and efficiency in this field substantially. Such an assistant could also level the differences between patients referred to healthcare professionals at different stages of their development process and/or with varying experience and/or expertise. Once a PD diagnosis is present, specialists in the PD field should treat these patients. In most instances, implementing AI in this industry will simply create more jobs, and result in an overall more efficient service. Further, new positions will be created from the programmers needed, and our desired future will also allow more researchers to contribute to this field. Indeed, such a pilot project could very well result in major publications in top journals such as Science and Nature (rare within psychotherapy research), and benefit both the involved organization, society, and patients.

As in most digital projects, the major challenge for implementing AI in this industry is the disconnection between the real world and it's digital representation. In this case the major challenge is not the lack of suitable data, but the process of being allowed to research on all historical records of psychiatric patients with PDs. However, as the benefits would be tremendous, and a deep learning algorithm would not need to keep track of the data it digests, but only remembers the high dimensional space with points of feature vectors, patient’s total anonymity could be ensured. As we shall see below, it is essential that the public be informed that there is always a "human in the loop", and that the human is in charge for all major therapeutic decisions. The goal of this project, and all future implementations, is the collaboration between humans and machines. Importantly, in this project, involving deep learning, humans could also learn from the maps the machines draw of our psychiatric categories – allowing our understanding of these concepts to broaden. If successful, the pilot could lead to further implementations within the field.

One current scenario from the debate around self-driving cars is that the brakes of an autonomous vehicle become inoperable. The car is heading toward a group of pedestrians who would be killed if the car hit them. The car has a choice to swerve and hit only one pedestrian rather than the group. Should the car swerve? Or, what if the car could swerve and avoid the pedestrians, but would thereby harm the occupant(s) in the car?

The advent of artificial intelligence (AI) is one of the defining opportunities for all organizations today. What is needed for an organization to rise to that opportunity and exploit the potential of AI at scale? In this current project, the answer is likely first and foremost information: Successful implementation of AI demands that leaders, clinicians, researchers, patients, and society learn to work and think in new ways. There are essential ethical challenges that need to be considered in the project. First of all, adding AI in this context changes the social contract (security, quality, efficiency) and may influence public trust. Hence, it is essential that the public is informed that there is always a “human in the loop”, and that the human is in charge for all major therapeutic decisions: The algorithm can only advice. Should the social contract be changed for the worse, this may even result in a much lower outcome, as we know that around 40% of clinical improvement stem from placebo, e.g., social expectations. This effect has been demonstrated to evaporate in the context of machines. Hence, information is crucial before, and at every step of implementation. One example, which highlights this, is the diagnostic and therapeutic evaluation of suicidal patients. If clinicians are influenced by machines in their decision not to hospitalize a given patients, it needs to be clear who made the decision and why. This is something that needs further investigation before implementation. Consequently, the first step in this project is seeking approval from the ethical committee, and discussing possible pitfalls, and how to best avoid them further down the road. Similar ethical concerns and dilemmas are currently being debated when it comes to autonomous vehicles. At issue is that machines would be making moral decisions. Such scenarios are typically unsolvable dilemmas, and different countries have different solutions to this problem. However, most agree that an autonomous vehicle should not be allowed to harm or kill its owner, as this would incentivize few people to buy such a machine. In terms of our project, there would never be a machine making a decision on it’s own, but the consequences of adding of a machine in the loop still needs careful consideration. Hence, ethical concerns will be less dramatic, and have milder consequences: It may for instance be that this ethical committee finds that as a general rule, machine learning should be used to assist the diagnosis of those diagnostic categories that benefit the most from tailored treatments. Next, this project demands that the public is informed through thorough press releases. It seems likely that this project would be so novel that journalists from all major channels and newspapers would be attracted, and the topic may well be debated public. Hence it will be important that the project has good communicators, who are able to inform the public that new technologies bring about new dilemmas, which have no clear-cut answer, but which needs to be tackled by society over time. Importantly, the presence of dilemmas should not in itself stop the technology, or the project, from proceeding, as we see from the current implementation of autonomous vehicles. The experts in the project would not need to answer all dilemmas themselves, but could also ask for experts from for instance MIT or Berkeley to help inform and guide the public opinion. As long as the project remains a pilot, the most important task is to get the work of informing the public going, such that dilemmas and issues can be debated in parallel with the implementation of the technology. Clinicians and researchers would need to be trained, and expert programmers would be hired to implement the ML algorithms.

Long-term, the implementation of AI in public mental care would likely not risk eliminating the need for clinical employees, as the waiting list in most developed countries are long. In this industry, it is crucial that there is a combination of humans and machines, allowing for clinicians to be more efficient and accurate in diagnosing and screening patients informed by AI. So, it seems that in most instances, implementing AI in this field will simply create more jobs, and result in a more efficient service. New positions will be created from the programmers needed, and our imagined future will also allow more researchers to contribute to this field. However, as the skill set needed to fulfill some of the positions where machines and humans need to cooperate may make the current work inappropriate for certain people, some employees could end up worse off after this technological change. However, if done skillfully, this should affect few people so negatively, and the majority of the organization would hopefully be better off. The observant reader has most likely already concluded that AI in mental healthcare may not be as mad as it sounds – but rather something we need to prepare for (perhaps even embrace?), as it seems both unavoidable (because of the benefits) and desirable (if done skillfully). I would totally agree.

P. S. I also think that there is a pandorean box to be opened somewhere in the future in terms of allowing AI to assist not only diagnosis, but also psychotherapy. This is not on the horizon yet, but would become relevant once algorithms are able to learn from stories (which is supposedly where we are headed).

Quiz

1. What is the main goal of using AI in mental healthcare?
   a. To replace human clinicians
   b. To improve the efficiency and precision of treatment
   c. To reduce the cost of healthcare
   d. To eliminate the need for human interaction

2. What is one potential application of AI in mental healthcare?
   a. Diagnosing and treating mental health disorders
   b. Providing personalized treatment plans for patients
   c. Tracking patient progress over time
   d. All of the above

3. What is a key challenge in implementing AI in mental healthcare?
   a. Ensuring data quality and diversity
   b. Ensuring patient privacy and confidentiality
   c. Addressing potential bias in AI algorithms
   d. All of the above

4. What is one way that AI could potentially improve mental healthcare?
   a. By providing more accurate and consistent diagnoses
   b. By identifying potential mental health issues earlier
   c. By personalizing treatment plans for individual patients
   d. All of the above

5. What is one ethical challenge associated with using AI in mental healthcare?
   a. Ensuring the social contract between patients and healthcare providers is not compromised
   b. Ensuring that human clinicians are responsible for all major therapeutic decisions
   c. Ensuring the use of AI does not diminish the importance of the human element in mental healthcare
   d. All of the above

6. What is one way to address potential bias in AI algorithms in mental healthcare?
   a. Ensuring that the data used to train the algorithms is diverse and representative
   b. Ensuring that the algorithms are tested on a wide range of patients
   c. Ensuring that the algorithms are transparent and explainable
   d. All of the above

7. What is the error rate for human radiologists in the study mentioned in the article?
   a. 0.5%
   b. 3.5%
   c. 7.5%
   d. 10%

8. What is the error rate for AI systems in the study mentioned in the article?
   a. 0.5%
   b. 3.5%
   c. 7.5%
   d. 10%

9. What is the error rate for the combination of humans and AI in the study mentioned in the article?
   a. 0.5%
   b. 3.5%
   c. 7.5%
   d. 10%

10. What is the main challenge for implementing AI in the field of mental healthcare?
    a. The lack of suitable data
    b. The process of gaining access to historical records of psychiatric patients
    c. The disconnection between the real world and its digital representation
    d. Ensuring the social contract between patients and healthcare providers is not compromised

Correct answers

Answer key: 1. b, 2. d, 3. d, 4. d, 5. d, 6. d, 7. b, 8. c, 9. a, 10. c