A Real-World Application of Machine Learning AI in Value-Based Care

A Real-World Application of Machine Learning AI in Value-Based Care

Dan Neff

VP, Engineering, qrcAnalytics

February 5, 2026

Perspective Matters

Having worked in healthcare software development since 1986, I’ve seen multiple technology waves come and go. From magnetic-tape–submitted claims and 1200-baud modems, to HL-7, open systems, clinical repositories, the web, and big data—each wave promised transformation. Some delivered. Some faded.

Artificial Intelligence is different.

AI is already proving meaningful value in healthcare, particularly at the point of care—imaging analysis, clinical decision support, and clinician productivity tools like dictation. These use cases are tangible, measurable, and improving quickly.

At qrcAnalytics, however, our focus is narrower and more pragmatic: value-based care, specifically quality measurement and HCC risk adjustment. That distinction is critical when evaluating where AI meaningfully applies.

Quality Measures vs. Risk Adjustment

Most quality measures are, by design, deterministic. A member either received a screening/service or they didn’t. AI does not currently add value in calculating a quality score itself. This does not apply to certain “risk-adjusted utilization measures” such as Plan All-Cause Readmissions (PCR), Emergency Department Utilization (EDU), or Acute Hospital Utilization (AHU).

HCC (Hierarchical Condition Category) risk adjustment is fundamentally different.

Risk adjustment is longitudinal, pattern-driven, and probabilistic. It depends on how diagnoses emerge and cluster over time, how patients interact with the healthcare system, and how conditions are documented. This makes HCC an ideal candidate for machine learning.

Machine Learning, Simply Stated

Machine learning is a subset of artificial intelligence that enables systems to learn from historical data and improve over time without explicit programming. Instead of hard-coded rules, models are trained on examples—diagnoses, encounters, and utilization patterns—and learn relationships that are difficult to define manually.

Our Approach Using ML.NET

Using Microsoft’s ML.NET library, we trained a machine-learning model with:

• 3 years of historical diagnosis data
• ~50,000 Medicare Advantage members
• Members with a CMS-HCC V28 captured in 2024

To preserve clinical accuracy/integrity, only diagnoses associated with Evaluation & Management (E/M) visits were included. “Rule-out” and lab-associated diagnoses were intentionally excluded.

The trained model was then applied to 2024 members without a V28 HCC capture, processing their historical diagnosis profiles to identify clinically likely—but undocumented—HCCs.

Model Outputs: Designed for Action

For each targeted HCC, the model generates two outputs:

Score

• A numeric value representing how strongly the model leans toward an outcome
• Not constrained to a 0–1 range
• Used internally to determine the predicted classification

Examples:

• Score 2.3 → strong positive signal
• Score -1.1 → negative signal

Probability

• A normalized value between 0 and 1
• Derived from the score (e.g., sigmoid transformation)
• Designed for operational decision-making

Examples:

• Probability 0.85 → high likelihood
• Probability 0.15 → low likelihood

In practice, probability is the primary value used by clinical and operational teams.

What the Model Uncovers

Two consistent categories of opportunity emerged:

Missed Recaptures

Conditions documented in prior years but not recaptured in the current year.

While qrcAnalytics already identifies missed recaptures using CMS risk models, machine learning demonstrated the ability to surface these opportunities without relying on CMS logic at all—purely through diagnosis patterns.

Common drivers include gaps in patient engagement, missed wellness visits, specialist-only care, and documentation variability.

Diagnosis-Pattern–Based Opportunities

More compelling are cases where diagnosis patterns suggest a clinically likely condition that has never been captured for the member.

Example

• Score: 0.17
• Probability: 54.24%
• Targeted HCC: CMS-HCC 267 (V28) – Deep Vein Thrombosis and Pulmonary Embolism
• Member: 72-year-old Female

Diagnosis Pattern:

DX Code	Description
E78.2	Mixed hyperlipidemia
I72.0	Aneurysm of carotid artery
I82.890	Acute embolism and thrombosis of other specified veins
I86.8	Varicose veins of other specified sites
N32.81	Overactive bladder
N39.0	Urinary tract infection, site not specified
R22.1	Localized swelling, mass and lump, neck
R59.0	Localized enlarged lymph nodes
Z01.810	Encounter for preprocedural cardiovascular examination

Based on ML analysis, there is a 54% chance this member has or may develop HCC 267 Deep Vein Thrombosis and Pulmonary Embolism. No single diagnosis confirms the condition. The pattern, however, supports clinical review—precisely the type of insight machine learning is well-suited to provide.

Productization and Roadmap

These CMS-HCC machine-learning libraries represent a natural extension of the qrcAnalytics platform.

Ongoing enhancements include:

• Continuous retraining with new member data
• Expanded feature sets including:
  • Age and sex
  • Vital signs
  • SNOMED-based clinical concepts

What’s Inside the Module

• Insights derived from 3+ years of longitudinal claims data
• High-precision identification of potentially undocumented but clinically plausible CMS-HCC V28 conditions

For provider groups and risk-bearing entities, this translates to:

• Improved risk capture
• Stronger coding accuracy
• Revenue integrity
• More complete representation of patient complexity

Call for Early Beta Sites

qrcAnalytics is engaging provider organizations, ACOs, and health plans to participate in early beta deployments.

Beta participants receive:

• Preferred pricing
• Early access to enhancements
• Direct influence on product direction