Managed care organizations serving Medicaid patients face a fundamental challenge: while traditional risk-scoring models excel at identifying patients likely to experience high costs, they don’t always identify which patients will actually benefit from specific interventions.

Take the following example: a patient who experiences a terrible car accident and is on a ventilator in an ICU is likely to face high costs, but is unlikely to benefit much from a phone call from a population health care manager. By contrast, a patient who has a history of falling down and was just started on a blood thinner might not have high costs yet, but a call from a clinical pharmacists to prevent over-dosing on the blood thinner, and from a care coordinator to get a walker to prevent the next fall, may prevent a critical head injury and months in the ICU with a brain bleed.

Research finds that prioritizing patients based on their likelihood to benefit from a given intervention, such as population health team outreach, consistently and successfully addresses this failure to consider preventive care windows of opportunity and prioritize who can best benefit from an intervention at a given time. Instead of prioritizing patients based on predicted costs , patients are prioritized based on which intervention will benefit them most at the current time, optimizing preventive care outreach. 

Challenges of Traditional Risk-Based Prioritization

Current population health team prioritization methods rely heavily on predictive models designed to identify patients most likely to experience high-cost events. These risk-scoring methodologies are statistically sound for actuarial purposes, but they make an assumption that isn’t always true: that patients with the highest predicted future costs will always derive the greatest benefit from population health outreach and interventions. This assumption doesn’t always yield the most accurate results, which can harm patients and hamper the quality of their outcomes.

A better question to ask would be, "Who is most likely to benefit from this specific intervention?" The relationship between baseline risk and treatment benefit varies substantially across patient populations, intervention types, and clinical contexts, creating systematic misallocation of care management resources.

This misalignment becomes particularly obvious in Medicaid populations, where social complexities lead to more variation in treatment responsiveness. The result is suboptimal resource allocation that undermines value-based care contract performance and perpetuates health disparities.

Published evidence demonstrates that risk-based prioritization directs intensive interventions toward patients who may be least responsive to standard care management approaches while overlooking patients with moderate baseline risk but high intervention responsiveness. For example, patients with advanced chronic conditions and complex social barriers may score extremely high on traditional risk models yet demonstrate limited responsiveness to telephonic care coordination. In the same vein, patients with moderate clinical complexity but specific behavioral health conditions might demonstrate exceptional responsiveness to integrated care management despite lower risk scores.

Traditional Models vs. HTE-Informed Approaches

Prioritizing patients based on their likelihood to benefit from an intervention represents the variability in treatment benefit across individual patients based on their comprehensive clinical, social, and demographic characteristics. Unlike traditional risk prediction, which focuses on identifying patients likely to experience adverse outcomes, heterogeneous treatment effects (HTE) methodology explicitly models the effectiveness of different interventions, and how that effectiveness might vary across different patient populations.

To see this in action, imagine two patients receiving Medicaid. One, a 34-year-old woman with diabetes and hypertension, scores moderately on a traditional risk prediction model (one scoring patient based on risk). This patient’s case isn’t going to trigger intensive care management; her medication adherence is decent and she hasn’t been hospitalized.

Standing beside her is a 68-year-old man with heart failure and COPD. He was hospitalized five times last year, and scores at the very tip of that same traditional risk prediction model. 

What happens next? The 34-year-old patient ends up in the ED multiple times over the next year due to diabetes complications that could’ve been prevented by targeted interventions. And, despite targeted interventions, the 68-year-old male patient is still hospitalized often because his condition requires management that population health team  coordination alone can’t address.

While the traditional risk prediction model correctly identified the latter as high-risk, it completely missed the former, who would have actually benefitted most from targeted interventions. An HTE-informed approach, which considers factors against what intervention a patient is most likely to benefit from, would have effectively caught both patients and ensured they both were connected to the quality care they needed to manage their respective risk factors.

Standard Risk Model Performance Limitations

Traditional Medicaid risk models, exemplified by the widely-used Chronic Illness and Disability Payment System (CDPS), demonstrate modest predictive performance when evaluated against clinically meaningful outcomes. CDPS, which relies on patient demographics, diagnostic codes, and medications to predict healthcare costs, achieved R² values ranging from 0.022 to 0.050 when predicting actual medical costs across different Medicaid population subgroups, meaning they explain only a small amount (2.2%-5%) of future cost variations. When applied to a dataset of 10 million patients receiving Medicaid benefits, we found that traditional approaches miss approximately 96% of patients who would benefit from proactive outreach, representing a substantial operational inefficiency for value-based care programs.

HTE-Based Model Performance Advantages

By contrast, advanced machine learning models incorporating HTE principles demonstrate dramatic performance improvements across all clinically relevant metrics. When applied to the same 10 million patient Medicaid dataset, HTE-informed models achieved a more than three-fold improvement in sensitivity (meaning the probability of correctly identifying a person in need) over traditional risk-based approaches.

Importantly, this improvement occurred without compromising the probability of truly identifying people who don’t need intervention, which remained at 99.8% compared to 99.5% for standard models. While this decimal shift may seem small, this performance characteristic addresses a key operational concern for care management programs: avoiding false positives that reduce outreach efficiency and increase program costs.

The discriminative ability–the ability to differentiate higher- from lower-benefit patietns–of HTE-informed models reached 79.5% for predicting non-emergent acute care visits, substantially exceeding the highest performance reported in previous Medicaid risk modeling literature: 67.7%. 

Addressing Bias Through HTE Methodology

CDPS models consistently underpredict costs for Black patients and overpredict for white patients, with differences ranging from $11-$46 per member per month. This bias reflects the well-documented phenomenon where Black patients have lower access to high-cost healthcare centers, resulting in lower observed costs despite similar or greater disease severity compared to white patients.

More significantly, HTE methodology reverses the lower sensitivity of risk prediction for Black versus white patients. This bias is reduced even when race and ethnicity is removed from the predictive model, which suggests that HTE approaches' incorporation of social determinants and advanced analytical methods is able to address structural factors that traditional models fail to capture. For MCOs operating under regulatory scrutiny regarding health equity, this performance characteristic offers a substantial competitive advantage.

Leveraging These Learnings to Improve Care Quality

Waymark Signal™,  our proprietary machine learning technology, demonstrates the practical application of these advantages in real-world Medicaid care management. Signal combines data on social risk factors and patient risk trajectories with healthcare utilization to identify patients at risk for preventable ER and hospital visits. That same identification also empowers community-based care teams to prioritize patients who are not just at high risk, but who are most likely to benefit from a specific intervention.

Signal’s incorporation of social determinants data addresses a critical limitation in traditional risk models that rely predominantly on clinical claims information. By integrating housing stability, food security, transportation access, and community resource availability data, Signal captures intervention response modifiers that clinical data alone cannot identify. As one of the largest and most representative comparisons of Medicaid risk models to date, published research in Nature Scientific Reports validates Signal's performance and provides concrete evidence of HTE performance over traditional risk-based approaches in actual care delivery environments.

For managed care organizations operating in value-based Medicaid, HTE methodology represents a much-needed transition from risk prediction to treatment benefit prediction. The results – measurable clinical improvements, improved care management effectiveness, and better value-based contract performance – speak to a future we can easily envision, one where all patients receiving Medicaid benefits can access the proactive, equitable care they need to live healthier and more productive lives.