Technology for Health Equity through SDoH & Health Data Integration

Under this thrust, we will merge, for example, SDoH information with clinical health records (e.g., medical history, vital signs, genomic, treatment and outcomes data) to better develop, train, and validate machine learning models for several diseases. This will lead to more precise and accurate predictions and treatment recommendations for underrepresented groups. In turn, these improved predictions and recommendations will achieve just and equitable clinical outcomes. In another set of projects, we aim to further integrate this information with mental health data (e.g., mood, behavioral patterns collected passively and actively with patients) to create comprehensive patient profiles, then leverage these profiles during treatment. While this aim will address a wide range of diseases currently studied at UI Health, as indicated above, such as diabetes, cardiovascular disease, or depression, below we describe, as two examples, planned approaches and outcomes in cancer treatment and in mental health.

Cancer Treatment:

How might an improved client (patient) profile generated through machine learning lead to better diagnostic and therapeutic outcomes? In the domain of cancer, one example might be genomics markers which have recently added to our knowledge of diagnostics and in understanding therapy responses. However, these first generation of markers have been trained on limited numbers of patients with a limited number of variables collected. Because of the relative homogeneity of participants (e.g., patients presenting to other large academic cancer centers) for both training and validation, certain algorithms have failed patients in the 'real-world.' The diversity of patients both in disease presentation and race and ethnicity and social/economic diversity has led to the underperformance of these algorithms. Take for instance the 21-gene recurrence score in breast cancer (RS). The RS is an expression assay that permits stratification of patients with genomic high risk hormone-receptor positive HER2 negative early stage breast cancer to be assigned chemotherapy to optimize survival and also permits patients with genomic low-risk breast tumors being de-escalated away from chemotherapy. This widely lauded test is now in routine clinical use and incorporated on national guidelines. However, emerging evidence from investigators at UIC has shown that in the real-world, this assay severely underperforms specifically for urban black women whose genomic risk underperforms for actual risk for recurrence, and steers women away from potentially life-saving chemotherapy. The test is reasonably precise but notably inaccurate for underrepresented Black women. By crafting a more holistic 'picture' of the client for clinical variables, genomic attributes, and for social factors that drive inequity but also capture effects of societal weathering (e.g., allostatic load), our institute will permit better derivation and validation of biomarkers. Put another way, advanced risk prediction models emphasizing inclusiveness, will be both precise and accurate for underrepresented groups to achieve equitable health outcomes for all groups.

Mental Health:

Mental health within underserved communities still has a stigma associated with discussion or seeking treatment. Novel methods to engage patients are warranted to include this critical dimension in the treatment of patients. We will first create Comprehensive Patient Profiles: AI tools can help merge SDoH information with clinical health records (e.g., medical history, vital signs) and with mental health data (e.g., mood, behavioral patterns) and social data to create comprehensive patient profiles. We aim to create 360-degree views of patients, integrating clinical, social and mental health data from diverse and varied sources. We will employ AI-based algorithms to analyze these comprehensive patient profiles, with a specific focus on addressing the unique healthcare needs of minority and marginalized populations, thus advancing health equity and ensuring improved access to quality healthcare for all individuals, regardless of their background or circumstances. Critical to the engagement of these tools is using terms and concepts that the community is comfortable engaging with. Rather than focusing and using in patient communication the DSM-5 terminology leveraged as part of EHR (DSM-5 is the standard classification of mental disorders used by mental health professionals in the United States), we will research terminologies more acceptable during patient communication to underserved communities (e.g., “sad” or “lonely” rather than “depressed”).

We further aim to develop Automated Detection and Personalized Intervention: AI, particularly large language models (LLMs) and Generative AI, may enable early recognition of mental health issues, and tailor personalized treatment plans based on SDoH individual data. We will leverage clinical data, behavioral data from smart phones and other platforms to detect and diagnose early signs of mental health issues such as depression, anxiety, suicide ideation and substance use. Furthermore, we will pursue Virtual Therapeutic Support and Stigma Reduction: AI-driven virtual assistants, powered by LLM can provide continuous therapeutic engagement, extend access to underserved populations, and reduce the stigma associated with seeking help. While LLMs hold great potential for mental health support, their potential limitations include data bias and hallucinations, overreliance on keywords that differ across different racial and ethnic groups, and lack of nuanced understanding of the contexts. Addressing these challenges we intend to design, develop and test the effectiveness of context-sensitive, AI-driven models, virtual assistants and chatbots for providing appropriate mental health interventions.

We also aim to create a Digital Behaviorome - a new Omics approach to brain and mental health: With the growth in the use of smartphones and wearable health devices to naturalistically, passively, and unobtrusively measure physical health in natural environments, we seek to extend this to the realms of brain health and mental health with the same passive and unobtrusive methodology. The digital behaviorome provides a platform to "close the loop" by providing real-time objective feedback to the users. Traditionally, neuropsychiatric assessments are conducted in-person and in-clinic and currently some assessments are moving remote; however they are intrusive and are subject to patient-survey fatigue even though these new methods can help improve care. By contrast, the digital behaviorome could revolutionize not only how brain health can be measured, but also how interventions can be delivered and decrease the burden to patients to enter data.