Contributors: Kira Radinsky, PhD, Oren Sarig, PhD, and Nathan Shapiro, MD
To learn more about Kira, Oren, and Nathan, click here.
Managing undiagnosed behavioral health conditions is a major population health issue and creates increased costs of hundreds of millions of dollars. Smart AI implementation can help health plans proactively identify individuals in need of appropriate screening and management leading to reduced cost of care and improved patient lives.
There is a growing awareness of behavioral health issues that face millions of Americans. Previously thought of as “invisible,” these issues and their impact are now at the forefront, thanks in part to the added stresses of COVID-19. Behavioral health (BH) conditions can lead to substance abuse, loss of income, increased likelihood of future chronic health issues, and can impair an individual’s ability to live a full, functioning life. More worrisome is the growing number of individuals whose behavioral health disorders remain undiagnosed. These untreated individuals suffer personally, are a burden on the medical system, and create billions of dollars in economic loss.
Out of the more than 56 million adults in the United States living with diagnosed behavioral health conditions, approximately 39.7 million did not receive treatment in the past year.1 Individuals with untreated behavioral health conditions account for $51.5 billion in workplace losses (productivity, short-term disability, and increased sick days)2. Depression (a leading untreated behavioral health condition) is the dominant medical disability for individuals aged 15-44, and the costs of untreated depression account for $26.1 billion in expenses to the healthcare industry and a total economic loss of over $83 billion annually.3
Identifying Undiagnosed Depression from the Data
Diagnostic Robotics, a leading digital health company has developed an artificial intelligence (AI) model that identifies patients with undiagnosed depression. This model assists health plans in identifying patients who are at risk for undiagnosed behavioral health conditions in their member pools. By relying on claims data points (the most common form of data available to health plans), the model uses deep learning technology to analyze each individual member’s personal medical profile, including past procedures, office visits, diagnoses, medications, and other indicators, to detect patterns of utilization indicative of an underlying behavioral health issue.
Reaching high precision with these machine learning models requires substantial under-the-hood technology. Medical claims need to be translated into features digestible by machine learning models. Diagnostic Robotics uses a hybrid clinical-AI approach whereby Diagnostic Robotics’ physicians provide the underlying medical insights about dimensions of the data worth exploring, and then a suite of machine learning techniques, including:
- embeddings (a technique used to transform the rich underlying data structure of medical records into a format that is digestible by machine learning algorithms, while preserving the informative content of the records)
- neural nets (also known as “deep learning”, an AI technique for creating prediction models that are not constrained by the simplistic mathematical nature of traditional statistical methods but can capture complex patterns of interaction between variables to deliver precise predictions), to slice and dice the underlying medical utilization data into accurate models.
The models also need to be audited for bias to promote healthcare equity. Without such a systematic review, certain disenfranchised populations may easily be under or over-represented in the predictions. The combined AI medical discovery platform can be directly applied to different payer data structures using a data warehouse structure mapping method.
Examples of Undiagnosed Depression Markers
As an example of signals picked up by the AI model, meet Patient 1, a 48 year-old female, who has undergone multiple musculoskeletal X-rays (including elbow, shoulder, knee, and spine). While multiple X-rays are not necessarily indicative of an underlying behavioral health issue, they are a marker in combination with other signals such as, behavioral health related diagnosis, utilization of core medical procedures, and additional chronic conditions. The AI model picked up on these disparate X-rays, in combination with other utilization patterns, as indicative of underlying behavioral health issues. Patient 1 was indeed diagnosed with major depressive disorder after the model prediction.
As another example, meet Patient 2, a 29 year-old male, with multiple past prescriptions for alprazolam (Xanax). The patient did not have a depression diagnosis when flagged by the model, but the model picked up on the chronicity of the prescriptions as a risk marker, which in combination with other factors in the patient’s medical background, flagged the member as high risk for undiagnosed depression. The member was eventually diagnosed with depression within two years (long after the model prediction) and began treatment with SSRIs.
Driving Clinical Outcomes and Cost of Care Improvements with AI
The model exhibits good performance in being able to identify the disease earlier than other modes of diagnosis and has the potential to improve patient outcomes by providing care earlier. Of the patients flagged as high risk for undiagnosed depression, 24%-52% (depending on the population) will be diagnosed with Major Depressive Disorder (MDD) over the following two years after the model’s early warning. This figure likely understates the actual prevalence of undiagnosed depression, since some of these members might never receive a diagnosis despite the presence of underlying medical conditions which are associated with or can lead to depression.
Insured individuals with undiagnosed BH conditions are also untreated (for their BH conditions); untreated individuals have a much higher cost of care. Among a population of insured individuals with COPD analyzed by Diagnostic Robotics, those with untreated BH conditions each cost their insurance plan $1,644 more in annual costs as compared to a matched sample4 of insured individuals with similar medical profiles.
Moreover, insured individuals with a high risk of undiagnosed BH conditions have a higher probability of becoming “cost bloomers” (patients who experience a surge in healthcare costs over multiple years). In fact, they are 2.8x more likely to become cost bloomers than the general population. When these individuals become cost bloomers, they become more expensive than average cost bloomers. Diagnostic Robotics’ models have found an average cost bloomer creates a $51,948 per member per year (PMPY) increase in costs, while an undiagnosed BH cost bloomer creates a $58,888 PMPY cost increase on average.
Implementing smart AI can help health plans reverse these troubling trends and avert a major population health crisis. By proactively identifying individuals at-risk, it can dramatically reduce the cost of care and improve countless patient lives.
Contact Kira at: [email protected]
Contact Oren at: [email protected]
Contact Nathan at: [email protected]
- Substance Abuse and Mental Health Services Administration, National Survey on Drug Use and Health, 2017. GAO-19-274.
- Greenberg P, et al. (2003). The economic burden of depression in the United States: How did it change between 1990 and 2000? Journal of Clinical Psychiatry, 64(12), 1465-1475.
- Leahy, R. (2010). The Cost of Depression. The Huffington Post.
- Matching controls for member age, gender, chronic conditions and SDOH characteristics.