Can Artificial Intelligence help in the treatment of depression?
A recently published study in JAMA Internal Medicine showed that clinical depression is poorly diagnosed and often goes untreated. A survey of 46,000 Americans screened for depression found that 8.4% of those interviewed had depression, but of those only 28.7% had received any treatment. Further, of those who were being treated for depression, only 29.9% of them screened positive for depression. Many people with conditions that were not actually depression were being treated with antidepressants.
Unlike other conditions, depression is diagnosed clinically, through in-person evaluation or through answering evidence-based questionnaires. There is no blood test or imaging test for depression, as there is for other chronic diseases such as diabetes, high cholesterol and heart disease. Additionally, more than 50% of patients with mental illness (including depression) use primary care providers as their sole point of contact with the health care system, making primary care the de facto mental health system. The prevalence of major depression is 2 to 3 times higher in primary care patients than in the overall population, largely because depressed patients use health care more frequently. What is more, depression impacts other chronic diseases, worsening morbidity and mortality in those other conditions.
Given that depression costs the United States some $40 billion annually, and given that the main place where depressed people contact the health care system is with their primary care clinician, and further, given that depression is under-recognized and under-treated, while conditions other than clinical depression are over-treated with antidepressant medications – it is clear that improved screening and tools made available to primary care physicians would have a large impact on access to treatment for depression, as well as on its cost and quality.
Can AI help?
Conceptually, applying deep learning in order to identify previously unknown relationships in medical data could help identify individuals at risk for major depression. By finding similarities between the individual at hand and patterns in broader medical data, personalized survey questions can then be surfaced to identify with more precision those who may suffer from depression. Further, such comparisons can help guide treatment approaches that might work best for a given individual, taking into consideration multiple factors such as co-existing medical conditions, medications, family history, genetics and responses to questionnaires (such as PHQ-9 surveys).
The challenge with using deep learning, of course, is that it needs a large amount of medical data to work with and develop pattern recognition from. Medical data remains dogged by fragmentation into institution-centered silos, though record sharing is starting to become more universal (which helps somewhat). In order to enable effective Artificial Intelligence (AI) to help with on-the-front-lines clinical medicine, such as primary care, this is inadequate. One needs aggregated data about a given patient – data from all the EHRs in settings where the patient has been seen, together with claims data, patient-entered and device-entered data, and genomic data where it exists. Further, one also needs a truly massive reservoir of data upon which pattern-recognition (which is really what AI is all about) can create comparisons and identify risks and recommendations. In clinical medicine, we often do this already (though may not be conscious of it) – we assess a given patient, see a particular pattern, and compare it to similar patterns from our experience. With AI, the comparison becomes much more powerful when it is made with the global data set far beyond anyone’s personal experience.
To do this, the data needs to be organized into a graph-like structure. It is beyond just amassing data together in traditional formats. It needs to be transformed into meaningfully defined objects and relationships, and the connections between them. Similar to Google’s organizing the Internet using a Knowledge Graph for rapid and context-aware searching, Flow Health is organizing medical knowledge into a Medical Knowledge Graph. Beyond just saying that “this patient has this diagnosis, these medications, and these lab values,” we are working to compare the individual to the broader population to make personalized predictions — “given these patterns of diagnoses, questionnaire responses, genetics, medications and demographics, the probability of this person developing condition x is y, and the most effective treatment approach is z.”
Help for the primary care setting
As we have seen, much of the identification and treatment of depression is in the primary care setting. Treatment of depression, particularly by a busy primary care clinic, is most frequently with medications (52% of patients were recommended medications), and psychotherapy alone is the least recommended option (4% of patients). However, antidepressant medications, as seen in the recent study, are often used for conditions that are not actually clinical depression.
The kinds of benefits that artificial intelligence (AI) can deliver to clinicians, especially primary care clinicians, are twofold: (1) by presenting personalized depression screening tools to patients at risk for depression based on the patterns in their medical data (clinical data, questionnaire data, genomic data and the like), these tools can improve the accuracy of diagnosis; and (2) by using pattern recognition that can analyze a given patient’s individual situation and compare it to patterns found in the Medical Knowledge Graph, AI can automatically suggest best-practice treatment recommendations. It may be that a given patient will respond better to medication A than to medication B, or maybe other non-pharmacologic or other combined approaches will work in this specific case. This is the promise of AI – making individualized recommendations based on the detailed complex situation of a specific person, driven by comparing their pattern to those of others who are very similar. The current treatment of “trial and error” can be replaced with something much more individualized. This is the goal of Precision Medicine.
How do we make this happen?
AI as an assistive tool is unprecedentedly powerful. The sorts of insights and decision support, where recommendations, authorization processes, and capture of quality metrics are done automatically in the background, can liberate burdensome workflows for practicing clinicians.
For consumers, there are two areas where this level of AI can be seen: (1) interacting with the health care system via (increasingly universal) portals, where viewing results and other EHR-based data, prescription requests, and communication with the practices already are taking place – these can be made much more powerful with individualized “what does it all mean?” content in those portals; and (2) consumer-facing communities where people with similar conditions can coach, help and support each other in curated, intelligence-assisted ways.
To get to this vision of a future state, we need to start with building a Medical Knowledge Graph populated by huge amounts of data. This is the base data “substrate” upon which AI algorithms can operate and identify relationships in the data, both for general medical research, and also as a background when used to match patterns in an individual’s data. In additional to the Knowledge Graph, an individual’s data needs to be amassed and aggregated from all its sources (or as many as can be gathered). That way the pattern of an individual’s medical data can be compared to the Knowledge Graph, and truly meaningful recommendations can emerge.
In the case of the identification and appropriate management of depression, this can be powerful. Patient-facing questions can be intelligently presented, and potential gaps in what is known about a particular patient can be highlighted. Gaps in the EHR can be filled in. If appropriate and indicated, even recommendations for additional testing, including genomic testing, can be presented to the clinician and the patient, so that individualized recommendations can be made. The right thing can be done the first time, rather than discovered by trial-and-error, which is so often the case without these tools.
This is what precision medicine will look like.