Since Dr. Atul Gawande’s famous article “Hot-Spotters,” care management has quickly become a rapidly adopted approach for medical cost containment among complex, high-utilization patients. While initial savings were rapidly seen, problems slowly emerged in this approach:
- Care management is labor intensive and costly;
- After addressing the most extreme outliers, programs saw diminishing returns;
- Retrospective structured data correlates less and less with future utilization and costs; and
- Member churn reduces available retrospective data for relevant predictions on a health plan’s current population.
Illustrating these challenges, the New England Journal of Medicine recently published a randomized controlled trial from that same “Hot-Spotters” program showing that the program produces no difference in reducing costs or readmissions than the control group which received no care management intervention.
These care management challenges can be solved by AI & NLP by:
- Augmenting care management to enhance efficiency and effectiveness;
- Utilize contemporary data in real-time for stronger predictive validity in identifying current members;
- Deploy machine learning and NLP to create models built on unstructured and structured data that deliver more accurate, timely and actionable information.
In Article 2 of our series, we showed how everything in payor care management begins with accurate member identification and matching to the right intervention to anticipate, mitigate and possibly prevent major medical events.
In this article, we will expand on this idea to encompass broader care management activities than preventative interventions such as diagnostics, medications, mental health, behavioral health and social determinants. We will discuss exactly how to build or expand AI applications to drive unprecedented care management ROI and create a strategic advantage for payors and integrated delivery networks (IDNs).
Key Enablers – Social Determinants of Health and Comprehensive Member Data
SDOH Are Just the Beginning
The social determinants of health (SDOH) framework (Figure 1, below) illustrate how members can receive the exact same medical procedures at the exact same quality but experience vastly different long-term health outcomes. SDOH changes the playbook for improving health outcomes from simply focusing on the healthcare delivery system to the member, their life and environment.
In current operations by companies like Socially Determined, they provide useful intelligence by connecting members to specific ZIP codes, characteristics and deep information. All these are useful features in a comprehensive AI and NLP models, however, by nature they’re naturally limited to generalizable center mass of the bell curve. In other words, social determinants will be useful but there will be a small proportion (2% to 10%) who associated neighborhood, economics, community or education don’t correspond in the normal way.
This is where NLP and AI are particularly important. NLP can interrogate current language from the members themselves or observations from their direct providers to deliver a more accurate and timely prediction. Importantly, AI can continuously update models in the field as these environments or members change over time. AI and NLP provide the benefits of specificity and dynamism that aren’t found in off-the-shelf SDOH solutions.
Member Medical Records – A New Source of Data
Health plans utilizing old claims data as the only source of truth for model development and patient-level predictions frequently will find their real-world results will fail to match their high statistical scores. Health plans access to member medical records are growing with ACOs, value-based contracts, EMR partnerships, vertical integration.
Access to contemporary patient record information enables stronger predictions for care management interventions. Temporal variations of unstructured medical record data describing complex diseases significantly impact the probability of mortality and all-cause 30-day readmissions (Tang et. al. 2018). You can now see how hamstrung health plans are when relying solely on historical claims data for member identification and population segmentation. When health plans can finally apply NLP & AI to medical record data in real time, their care management program will deliver dramatically improved results.
Comprehensive Member Data
Going beyond structured historical claims data and generalized SDOH to ingest, analyze and deeply understand members will result in far more precise member predictions, risk stratification and, ultimately, cost containment and improved member health. In the table below, a summary of the new sources of data available for health plans and the new, actionable information that can be derived from applying NLP & AI.
Table 1. New Data, Actionable Insights & Results from AI & NLP Payor Intelligence
|New Data Source||Actionable Insights from AI & NLP||Results|
|Post-Discharge Phone Calls||Activities of daily living (ADL), safety event risk (e.g., falls, infections), medications||Reduce utilization of SNF, ED|
Reduce all-cause readmissions
Improve medication/therapy adherence
|Utilization Review Phone Calls||Denial / Approval Probability, Decision Reasons, Requirements, Evidence||Improve time to decision, Reduce clinical review, Increase UM policy adherence, Medical cost containment|
|Telemedicine Visits||Augment claims documentation, identify comorbidities, identify risks, identify members for preventive health screenings||Improve claims processing time, reduce clinical review, improve UM policy adherence, medical cost containment, earlier diagnoses and reduce misdiagnosis|
|Member Calls to Customer Service or Benefits Navigators||Identification of social isolation, PTSD, suicidal ideation, recent loss of spouse/partner and other behavioral health issues||Identify and address potential mental and behavioral health comorbidities, Improve member health and contain medical costs|
|Care Management Calls & Notes||Therapy and medication adherence, ADL, comorbidities, UM policy application||Improve adherence, Identify and address potential mental and behavioral health comorbidities, improve member health and contain medical costs|
|Medication Review||Polypharmacy, medication errors, adverse events||Patient safety, member health, medical cost containment|
|Member Medical Records||Comorbidities, disease risk, mental/behavioral health risk, medication risk, care coordination, future medical events, missed or delayed diagnoses, clinical follow-up, screenings/preventive health||Improve time to decision, Reduce clinical review, Increase UM policy adherence, Medical cost containment, Improve time from diagnosis to treatment, Reduce misdiagnosis|
Mental & Behavioral Health Drives Medical Healthcare Costs
You’ll notice that a major theme in the growing application of NLP & AI (Table 1) will be the identification, triage and care management for mental and behavioral health issues. Growing evidence shows that medical costs — especially among members with multiple chronic conditions — are frequently driven by or find their root cause in mental or behavioral health.
In the past, proactively identifying, diagnosing and treating the mental and behavioral health needs of a population was neither cost effective nor possible given the limited providers and clinical delivery model. Augmented with AI, telemedicine’s mental and behavioral health services solve the challenges in capacity, access and costs which previously limited their viability.
Ultimately, care management ROI with AI & NLP comes from three core areas: (1) enhanced, real-time and highly accurate identification of members’ needs, (2) augmenting the care management team to create strong efficiencies, and (3) expanding the features and complexities of models to include the widest range of potential causal factors — including mental and behavioral health.
All this will be predicted upon an NLP & AI with a comprehensive ingestion engine, capable of consuming and reasoning over all types of data — voice calls, audio transcripts, care management notes, and even entire medical records from multiple provider sources.
Using AI to Turn Care Management Into a Strategic Advantage
Results from Holistic AI Clinical Surveillance
Misdiagnoses are a leading cause of medical errors affecting a large population and causing serious medical harm across every care setting. In U.S. Hospitals, 40,000 to 80,000 deaths occur each year involving a misdiagnosis, and, in total, approximately 12 million Americans suffer a diagnostic error each year in a primary care setting—33% of which result in serious or permanent damage or death (Newman-Toker et. al. 2019). In emergency medicine, 12% of patients are misdiagnosed with an associated increased mortality rate (Hautz et. al. 2019).
Increased Preventative Health
AI & NLP are now being applied to increase the appropriateness and frequency of effective preventative health measures — particularly cancer screenings. The standard “age, sex and history” heuristic of determing who and when someone should receive a screening leaves out an unacceptable portion of the population. With today’s computing technology, the comprehensive risk factors can be automatically calculated enabling proactive outreach, scheduling and screening with relatively low manual effort — and high efficacy.
Improved Medication Adherence
Medication adherence has been a problem for decades. Claims data is not very effective in predicting future medication adherence (Zullig 2019) nor is purchasing information (Krumme et. al. 2015). The only structured variable consistently predictive of medication nonadherence is past medication nonadherence. To identify members who will experience difficulty in medication adherence, payors need to go beyond quantitative data by analyzing qualitative, unstructured data such as previous medication adherence (Kumamaru 2018).
Partnering with Providers
While all of these solutions and results are enabled by AI & NLP, it’s critical for health plans to understand their provider networks and effectively partner with them to deliver optimal care. In the next article, we’ll discuss how AI & NLP can be applied to better understand provider behavior and utilize deep profiles to make significant improvements.