Health Equity and SDOH
Bridging the Gap Leveraging Heatlh Information Technology to Address Healthcare Inequities
- This poster presentation examined disparities with a focus on The R4P (Remove, Repair, Restructure, Remediate) tool for designing equitable HIT solutions.
- The presentation highlighted 2 case studies. The first demonstrating bias in pulse oximetry due to higher readings on patients with darker skin tones, which in turn leads to delays in care. The case study supports the need for redesigning devices with diverse skin tone datasets, and AI driven corrections. The second case study highlighted disparities in radiology care and supported AI assisted scheduling, mobile radiology units, and improved insurance options to reduce the disparity.
Building Healthier Communities Strategies for Prioritizing Equity and Culturally Competent Care
- This presentation reviewed the importance of addressing equity and social risk factors, strategies to improve outcomes and reviewed the process of applying for Healthcare Equity Certification from the Joint Commission.
- The presentation showed the impact of broad approaches compared to targeted approaches for increasing breast cancer screening and lunch cancer screening.
Automated Extraction of Social Determinants of Health To Improve Patient Outcomes
- The presenter is in the process of developing an AI tool that reviews patient charts and extracts social determinants of health data from non-standardized fields (ie free text). The tool then can convert unstructured SDOH data into structure data fields.
- Currently, they are testing pattern matching approaches, Bidirectional Encoder Representations from Transformers (BERT) based models, and Large Language Models. Then working with the results to create predictive graphs for 30-day unplanned readmissions.
SDOH Program Design Technology Data and Evaluation that Drives ROI
- ROI needs to consider cost savings, efficiency, quality, satisfaction, and insights gained from programs.
- The presentation reviewed 2 case studies where health systems used Unite Us to address social needs:
- North Carolina Health Opportunities Pilots: program testing non-medical interventions for high needs Medicaid enrollees. Services include housing, food, transportation and toxic stress interventions. Participation led to an estimated $85 PMPM in Medicaid savings.
- Sarasota Memorial Health Care System/First 1,000 Days: The program included a community collaboration to improve care access for pregnant mothers and families with young children. The program connected participants to social care services. Saw a reduction in postpartum-related and all-cause 30-day readmission for Medicaid/Medicare patients.
Clinical and Community Data Initiative Data Linking for Food Security.
- Food insecurity significantly raises healthcare costs and ED visits.
- Nutrition services help reduce hospitalization and improve chronic disease outcomes.
- Clinical providers and community-based organizations operate in disconnected systems and CBOs cannot share or access clinical data due to technical and regulatory barriers.
- The goal of the project was to establish a common data model for clinical and social data, enable cross organizational data sharing, measure outcomes, and build dashboards for reporting and analysis.
- Outcomes:
- About 75% of CBO participants were successfully matched with clinical data.
- Roughly 45% of matched patients had diabetes
- A1C control improved modestly in 3 and 6 months after service delivery.
Artificial Intelligence
AI in Healthcare Avoiding Pitfalls and Driving Project Success
- AI projects are only as good as the data used for the model.
- Data quality and quantity both matter. Proof of concepts often don’t prove anything because the data does not reflect real world variables.
Las Vegas Cardiff Project AI Models Map Violence and Overdoses
- The Cardiff Model of Violence Prevention focuses on ER data providing solutions.
- One study in two police jurisdictions found that 83% and 93% of violent injuries seen in the ER were not reported to law enforcement.
- Using ER data, participants created a heat map of violent crime then using AI models mapped probability of violent crime and overdoses by location. Using the map, participants developed community driven interventions in the predicted crime hotspots. Interventions included Narcan training and street pastors.
Developing a Solid Foundation for AI Governance in Healthcare Organizations
- 45% of those surveyed through HIMSS are using AI/Machine Learning and 55% are not. Reasons for not using AI tools include:
- Lack of funding to purchase technology (45%)
- Technology does not fit in workflow (41%)
- Lack of organization policy/governance framework to implement technology (39%)
- Lack of and/or perceived lack of accuracy (27%)
- Ethical concerns about development and use (20%)
- Mass General Brigham shared their governance structure which includes:
- AI Governance Committee
- AI Implementation Oversight Working Group
- Individual Project Teams
- Slide 17 shows their phased approach to implementing AI.
- AI tools in use include using AI to summarize patient visits into clinical documentation and basket draft messages.
- Getting started: work to solve a particular problem, make sure everyone agrees on the terminology and technology, create an inventory of AI tools in use.
- Types of lability for Health AI
- Privacy
- Consumer protection and non-discrimination
- Tort Liability
- Medical malpractice
- Institutional liability
- Direct liability
- Vicarious liability
- Emerging Theories of Liability
- AI personhood
- Common enterprise liability
- Products Liability:
- Defects
- Failure to warn
- Strict malfunction
- Breach of warranty
- FDA approval and manufacturer liability
- ONC Certification Program
- False Claims Act
AIML Driven Clustering of Diabetes and Hypertension Populations
- The presenters conducted a retrospective analysis of patients with diabetes and hypertension over three years. The clustering analysis was used to identify patient groups based on social and clinical characteristics.
- 8 distinct clusters emerged from the analysis, two of which were identified as vulnerable (Latina with language barriers and middle aged, black males with high social risk).
- The two vulnerable clusters had the highest ED visit rate, lowest outpatient visit rates, and lowest digital engagement.
- 25% of patients with diabetes and 11% of patients with hypertension had not had a visit in the past 12 months.
- Using this data the health system will create targeted interventions to address care opportunities and engage patients in services.
Generative AI Security Essentials
- GenAI can be weaponized by cybercriminals:
- Phishing: highly targeted, scalable attacks.
- Identity Theft: deepfakes, voice cloning.
- Exploitation: reputational and financial harm.
- Disinformation: fake but convincing media content.
- Top Gen AI User Risks
- Security & Privacy: AI tools may learn from your input and leak sensitive data.
- Bias & Fairness: Models can reflect and amplify societal biases.
- Overreliance: Users may trust inaccurate or overly confident AI outputs.
- Miscommunication: AI struggles with nuance and can misinterpret vague prompts.
- Risk mitigation strategies:
- Do not enter sensitive data into public AI systems
- Audit AI output for bias and errors
- Cross check results with subject matter experts
- Establish AI use policies for ethics, security, and transparency.
- Organization AI Policies define approved tools, acceptable uses, data security practices, and transparency requirements.
Virtual Care
Unlocking Virtual Care A Collaborative Approach to Expansion and Success
- Valley Health (VA) shared their experience scaling and maturing their virtual care program. Their goal was to create a mature digital health strategy that is best in class, consistent, effective, and prepared to serve multiple clinical programs simultaneously.
- Valley Health created a steering committee to guide the rollout of services and developed a scorecard for their goals. They had one common virtual platform (previously had multiple platforms across their health system). Ambulatory virtual health services included virtual medical visits, school visits, community paramedicine and virtual urgent care.
- Starting on Slide 16, there are images of their equipment and set up for different services.
Interoperability
Pioneering Bidirectional Data Exchange for Mandatory Compliance and Beyond
- The presentation reviewed CMS Interoperability rules and challenges with payer-to-payer data exchange. Requirements include the ability to:
- Retrieve data from a prior health plan for all new members and at current member request
- Exchange data via FHIR API
- Integration of data into a longitudinal health record
HL7 AI Standards Provenance Fraud detection Prevention and Health Equity
- The presentation reviewed HL7 FHIR standards, AI in healthcare, collaboration for health equity, and addressing bias in AI.
- FHIR has evolved since its first proposal in 2011, with significant milestones in 2014, 2018, and upcoming releases. The FHIR community includes implementers, standards developers, and various stakeholders working towards interoperability and healthcare improvements.
- The FHIR Accelerator Program supports implementers in creating guides for public-private sector solutions in healthcare.
- Da Vinci Project: A multi-stakeholder initiative focused on value-based care, improving prior authorization processes, and enhancing clinical data sharing.
- CodeX Initiative: CodeX aims to improve cancer care and research through FHIR-enabled workflows, expanding to cardiology and genomics.
- FAST Accelerator: The FAST initiative identifies scalability gaps in FHIR resources and proposes solutions to accelerate FHIR adoption.
- Vulcan Project: Vulcan connects clinical research and patient care, focusing on standardized data exchange and improving clinical trial outcomes.
- Ethical concerns include bias, interoperability, and the balance of risk and reward in AI applications in healthcare.
- AI developers should ensure transparency and establish guidelines for monitoring bias in algorithms to promote equity.
- Strategies include partnering with equity organizations, ensuring accessibility, and prioritizing tools that address health gaps.
Cybersecurity
Proactive and Reactive Strategies to Minimize Data Disruptions in Healthcare
- The presentation discusses strategies to minimize data disruptions in healthcare through proactive and reactive governance measures.
- Challenge of Healthcare Data
- American hospitals generate an average of 2600 terabytes of data daily.
- 98% of healthcare leaders prioritize improving data quality to achieve organizational goals.
- Poor-quality data affects various areas, including scheduling, EHR, enrollment, and claims.
- Proactive and Reactive Strategies
- Proactive measures include planning, preventing issues, and understanding data needs.
- Reactive measures focus on detecting problems, containing issues, and restoring data integrity.
- Strategies include automating tasks, clear documentation, and defining expected input/output.
- Opportunities for Generative AI
- Generative AI can assist in anomaly detection and validating data accuracy.
- It can fill in missing data and promote consistency across systems.
- The effectiveness of AI depends on the quality of training data.
Comments