Predictive Analytics for Aging

The Medicare predictive analytics market is a crowded space where the distance between vendor claims and clinical evidence is rarely examined with precision. Every major population health platform claims the ability to identify the patients most likely to be hospitalized next month, stop filling their prescriptions, or fall within a 90-day window. Some of those claims rest on rigorously validated models with published performance data. Many rest on internally generated benchmarks, single-organization pilot results, or model metrics that measure training-set performance rather than prospective accuracy in live clinical deployment. The distinction matters because organizations making care management investment decisions based on risk scores are deploying real clinical labor — care coordinators, social workers, pharmacists — on the basis of those predictions. A model that fires on 30 percent of a panel because its threshold is set for sensitivity rather than specificity is not a clinical asset. It is an alert fatigue generator.

This article examines the evidence base across the three prediction domains most relevant to Medicare: hospitalization, medication adherence, and fall risk. It maps the named company landscape and the structural advantages that differentiate the platforms likely to matter in five years. It closes with the question payment models are actually asking: not whether the prediction is accurate, but whether the prediction generates an intervention, and whether the intervention changes anything.

What the Models Actually Get Right

#

Hospitalization prediction is the most developed and most extensively evaluated prediction domain in Medicare analytics. The published literature on hospital readmission models is large, and the performance of machine learning models against the clinical scoring tools they have largely displaced tells a consistent story: ML approaches produce modest but real improvements over simpler instruments in discriminating high-risk patients, but the improvement in AUROC from, say, 0.71 with LACE to 0.76 with a gradient boosted model trained on the same population does not translate into proportionally better clinical outcomes unless the workflow that acts on the score is redesigned to match. The best-performing condition-specific models — CHF readmission prediction, COPD exacerbation models, diabetes hospitalization risk — outperform general hospitalization risk models because they incorporate condition-specific clinical variables that general models treat as noise. A CHF readmission model that includes weight change, diuretic adherence, and sodium intake proxy variables derived from claims and pharmacy data has a structurally different feature set than a general readmission model using diagnosis codes and prior utilization.

The claims-based limitation is the binding constraint on most Medicare hospitalization models. Medicare claims capture what happened: the diagnoses, procedures, and encounters that generated a billing event. They do not capture medication adherence outside of what Part D prescription fills indicate, functional status, social support, housing stability, or the dozens of non-billable clinical observations that an experienced clinician uses to assess whether a patient is heading toward a hospitalization. Models that incorporate clinical data from EHR systems — vital sign trends, lab trajectories, nursing assessment flags — consistently outperform claims-only models for the populations where EHR data is available. The challenge is that the patients most at risk of hospitalization are often the patients with the least consistent EHR data, because they are the patients who do not see their primary care providers regularly enough to generate the clinical data the model needs.

Medication adherence prediction presents a different problem. Pharmacy claims data measures prescription fills, not pill ingestion. Proportion of Days Covered, the standard claims-based adherence metric, captures whether a beneficiary filled enough prescriptions to cover their days of therapy. It does not capture whether they took the medication, took it correctly, or took it on schedule. A beneficiary who fills 90-day supplies every 90 days has a PDC of 1.0 and may still be taking the medication inconsistently or not at all. Predictive models built on PDC inherit the limitations of the metric.

Social determinants are consistently among the strongest predictors of medication non-adherence in research settings, but social determinants data is the hardest data to obtain at Medicare population scale. Food insecurity, housing instability, transportation barriers, and social isolation all predict adherence behavior better than most claims-based variables, but they require either patient self-report or inference from non-clinical data sources. The organizations that have made the most progress on SDOH-integrated adherence models are those with direct patient engagement infrastructure — programs that generate self-reported SDOH data through screening questionnaires administered at clinical encounters or through community health worker visits. The data availability problem is not technical. It is operational: getting the SDOH data requires a clinical workflow that generates it.

Fall risk prediction was covered in the ambient intelligence context in MCR-06.08, but the claims-based dimension warrants distinct treatment here. Claims-based fall risk models — using prior fall diagnoses, emergency department visits for fall-related injuries, medication classes associated with fall risk, and functional limitation codes — produce useful population-level risk stratification but limited individual-level predictive precision. The signal-to-noise ratio in claims for fall risk is low enough that most claims-based models identify a high-risk cohort that includes many patients who will not fall and misses a meaningful fraction of those who will.

Passive monitoring data, as described in MCR-06.08, produces materially better fall risk prediction because gait parameters are leading indicators rather than lagging indicators. A claims code for a prior fall records an event that has already occurred. A gait speed deterioration signal from a continuous monitoring system records a physiological change that precedes the event. The most capable fall risk models combine passive monitoring gait data, medication burden data, and baseline clinical assessments, and the best-performing models in research settings achieve AUROC values around 0.76 for prospective 4-week fall prediction in dementia populations. That is a useful but not definitive signal — useful enough to prioritize clinical outreach, not reliable enough to treat as diagnostic.

The Named Company Landscape

#

Arcadia is the attributed population analytics platform most widely used by MSSP ACOs and ACO REACH entities. The platform aggregates claims, clinical, and pharmacy data, applies risk stratification, identifies quality gaps, and generates care management workflow tools. Nordic Capital acquired Arcadia in July 2025, accelerating its growth investment. Arcadia serves more than 30 percent of Newsweek’s 2024 Best Hospitals and has positioned itself as the analytics infrastructure layer for organizations with high-volume Medicare, Medicaid, and commercial value-based care populations. Its competitive positioning is data aggregation depth and clinical-financial integration: the ability to connect a risk score to a quality gap to a billing implication within a single platform.

Innovaccer occupies similar territory but with a stronger integration-layer positioning. The platform has deep WISeR vendor relationships, and its Ohio deployment as a WISeR prior authorization vendor gives it a federal contract data access foundation that creates structural advantages for Medicare analytics. Innovaccer’s Care Management Copilot uses LLM-based automation to generate care insights and documentation, combining population-level risk stratification with point-of-care clinical decision support. KLAS Research has recognized Innovaccer’s predictive analytics dashboard capabilities and its willingness to build custom solutions for large health system customers.

Lightbeam Health Solutions has established itself specifically within the ACO market. In performance year 2022, Lightbeam’s ACO clients managed care for 1.1 million patients across MSSP and REACH programs. The company’s November 2025 partnership with Wakely Consulting Group produced two ACO-specific tools: ACO Optimization Retrosight for retrospective performance benchmarking and ACO Optimization Futuresight for predictive network scenario modeling. KLAS has recognized Lightbeam’s deviceless RPM product — which integrates claims-based risk stratification with remote monitoring workflow — as a top performer in its category. Lightbeam’s SDOH Individual AI solution proactively identifies social vulnerability risks for attributed patient populations, addressing the data availability problem for social determinants by using algorithmic inference from available claims and demographic data where direct SDOH data is absent.

Health Catalyst serves health systems navigating value-based care transitions with analytics infrastructure. The company launched Ignite Spark in April 2025, providing enterprise-level analytics for community-based and ambulatory care settings. Health Catalyst’s positioning is more health-system-centric than ACO-specialist, making it a natural choice for integrated delivery networks taking on global budget risk through AHEAD or MSSP participation rather than for standalone ACO organizations.

Navina occupies a different segment: point-of-care patient summaries that synthesize claims and clinical data to support risk capture at the individual encounter level. The clinical application is primarily encounter-based risk adjustment documentation. A Navina summary surfaces HCC-relevant diagnoses, identifies coding gaps, and presents the clinical context supporting accurate documentation in the time available during a clinical visit. The encounter-based RA transition that CMS has been pursuing — shifting from retrospective chart review to encounter data submission as the mechanism for risk score calculation — makes tools like Navina more strategically important as plan-level RAF optimization through retrospective reviews faces tighter scrutiny.

Regard automates HCC documentation support within clinical workflow, identifying conditions present in the clinical record that qualify for HCC coding and generating draft documentation for clinician review. The compliance risk with tools in this category is the distinction between supporting legitimate clinical documentation of conditions that are present and managed, and optimizing coding for payment without corresponding clinical management. CMS’s aggressive posture toward chart reviews and retrospective coding creates regulatory exposure for organizations that use analytics tools primarily as coding optimization instruments rather than as clinical documentation supports.

What Payment Models Actually Buy

#

The demand structure for predictive analytics in Medicare is driven by risk exposure, not clinical preference. Fee-for-service Medicare creates no financial incentive for hospitalization prediction, because avoiding a hospitalization in FFS means forgoing the Part A payment for that admission. The more accurate the prediction and the more effective the intervention, the more revenue the fee-for-service provider sacrifices. This is not a technology problem. It is a payment model problem.

Two-sided risk ACOs in MSSP ENHANCED track and ACO REACH create the primary demand signal. An ACO with downside financial exposure has a direct financial incentive to invest in predictive infrastructure that identifies high-risk patients and deploys care management resources to prevent avoidable utilization. The ROI calculation is specific: if a care management intervention costs $800 per patient per year and prevents a hospitalization that would have generated $12,000 in shared savings liability, the model justifies aggressive investment in the analytics infrastructure that identifies the right patients for intervention. This is why MSSP ENHANCED and ACO REACH participants are the primary buyers of population health analytics platforms.

AHEAD global budgets extend the same logic to hospitals. A hospital accountable for total population spending has the same incentive structure as a two-sided ACO, amplified by the scale of the global budget: the hospital’s entire Medicare revenue base is within the budget envelope, not just the shared savings from an ACO population. This creates demand for analytics infrastructure that operates at health system scale rather than ACO scale.

MA plans use predictive analytics primarily for care management in high-cost member populations and for risk adjustment optimization. CMS’s escalating scrutiny of MA overpayment — including the chart review investigations described in MCR-02.02 — has changed the risk calculus for using analytics tools primarily for risk score inflation rather than clinical care management. Plans that built analytics infrastructure primarily around retrospective coding reviews face the same regulatory trajectory as the unlinked chart review practices CMS has been targeting. Plans that built analytics infrastructure around genuine care management interventions for high-risk members are in a more defensible position.

The Last-Mile Problem

#

The consistent finding across analytics implementations in health systems and ACOs is that model accuracy is not the binding constraint on outcome improvement. The binding constraint is whether the model’s output produces a clinical action, and whether that action is something the care team can execute with the resources they have.

An ACO care manager who receives a daily list of 45 high-risk patients flagged by the hospitalization prediction model, with no additional information about what specifically is driving the risk score or what intervention is indicated, will triage that list based on whoever responds to their phone calls. The model’s discriminating power does not improve outcomes if it is not connected to a workflow that specifies what action is appropriate for what risk signal. The implementations that have demonstrated clinical outcome improvement from predictive analytics have systematically invested in the workflow redesign alongside the technology: defining what the care team does when the alert fires, building the intervention into the schedule of the clinician who can execute it, and creating feedback loops that allow the model to learn from whether the intervention worked.

The Medicare market specifically requires one additional capability that general-purpose predictive analytics does not supply by default: Medicare-specific coding and billing rules embedded in the risk stratification and care gap logic. A general population health model trained on commercial claims data will misrepresent the risk landscape of a Medicare population, where HCC coding, Part D utilization patterns, the PDGM grouping logic for home health, and the SNF qualifying criteria all shape the data in ways that commercial-population models do not account for. The vendors with genuine Medicare specialization — built from Medicare data, validated on Medicare populations, calibrated to Medicare coding conventions — have a structural advantage over those applying general-purpose models to a Medicare context.

Related Reading

#

MCR-01_08 AHEAD and Geo AHEAD: Geography as a Cost Control Lever MCR-12_04 The HealthTech Company Ecosystem: What Medicare Policy Actually Allows vs. What Companies Claim