From CKD Risk Scores to Action: How Next-Best-Action Models Can Prioritize Kidney Care

Finding a patient at high risk for chronic kidney disease progression is an achievement. It is not yet an intervention. That distinction defines the next operational challenge in kidney population health. Once a risk model identifies a person with advanced or rapidly progressing CKD, the care team may uncover a daunting list of reasonable actions: arrange a nephrology visit, repeat laboratory tests, address uncontrolled blood pressure or diabetes, reconcile medications, remove an access barrier, or reconnect the patient with primary care. Each item may be important. Few patients or clinicians can act on all of them at once.

In this episode of The Strategy of Health, host Cole Lyons speaks with Joseph “Joe” Vattamattam, co-founder and president of Healthmap Solutions, and Stephanie Toth-Manikowski, MD, MHS, FASN, Healthmap’s chief nephrology officer, about the logic behind the company’s “next best action” approach. Their premise is simple: the value of prediction depends on whether an organization can translate it into a clinically meaningful, achievable next step.

The conversation builds on AJHCS’s prior coverage of proactive CKD population health management and the strategic shift toward earlier-stage kidney care. This time, the focus moves one layer deeper—from identifying risk to operationalizing it.

Risk stratification answers “who.” Care prioritization answers “what now?”

Traditional risk stratification helps a healthcare organization identify patients who are most likely to progress, experience an adverse event, or generate avoidable utilization. That information can focus scarce resources. But a high-risk label can also expose a large inventory of unmet needs without indicating which one should lead. Vattamattam described Healthmap’s early experience with this problem. A person living with CKD may also be managing diabetes, hypertension, cardiovascular or pulmonary disease, and multiple medications. Applying every relevant evidence-based guideline at once can produce a plan that is technically comprehensive but practically overwhelming.

“If you throw the kitchen sink at the patients, it’s overwhelming,”
The same burden compounds for providers managing dozens or hundreds of patients with limited time.

Healthmap’s response was to move from a catalog of possible interventions toward a ranked recommendation for an individual at a particular point in the disease journey. In the company’s formulation, the next best action is the care opportunity expected to have the greatest clinical impact for that member now—not a generic top-five list applied across the population.

That distinction is consistent with the direction of the wider kidney-care field. A recent Kidney Disease: Improving Global Outcomes discussion document asks whether clinical decision-support systems can recommend next-best actions from validated guidelines while minimizing alert fatigue and preserving clinician autonomy. The question reflects an important shift: the design challenge is no longer only whether a system can generate an alert, but whether it can produce an alert worth acting on.

What a next best action can look like in CKD care

Toth-Manikowski illustrated the problem with a patient from her own nephrology clinic: a man approaching stage 4 CKD whose kidney function had been declining rapidly and who had not seen her for two years. His diabetes and blood pressure were also poorly controlled. Cases like this are not reducible to a single disease marker. The care team must consider the stage and trajectory of kidney disease, the recency of primary and specialty care, missing laboratory data, medication opportunities, comorbid conditions, and the people or services with which the patient is already engaged.

Depending on the patient, a high-priority action might include:

• reconnecting someone with a nephrologist or primary care clinician
• obtaining missing kidney-function or albuminuria testing
• prompting a clinician to review a medication opportunity or safety concern
• addressing a barrier such as transportation, affordability, or behavioral health; or
• arranging follow-up after a signal of accelerated disease progression.

The 2025 NCQA white paper on CKD care gaps similarly emphasizes a holistic approach spanning risk-factor management, case finding, testing, medication access and adherence, education, and coordination. The practical implication is that care-gap closure is not a single workflow. It is a prioritization problem across several workflows.

For Toth-Manikowski, the question guiding the model was: If one member were selected from the population, what is the most clinically impactful thing the organization could do to help get that person back on the right path? That framing creates discipline. It does not imply that lower-ranked needs disappear. It sequences them so that the patient and care team have a feasible place to begin.

Designing clinical decision support clinicians will use

When decision support interrupts a workflow, lacks clinical context, or produces too many low-value prompts, clinicians learn to dismiss it. Next-best-action models therefore face an adoption test as important as their technical validation: Does the recommendation make sense to the person expected to act?

Vattamattam and Toth-Manikowski said Healthmap addressed that question by placing clinical, AI and machine-learning, analytics, and medical-economics teams in the same development process. Clinicians helped formulate hypotheses, interpret statistical findings, and test whether a recommendation would be appropriate for a primary care physician, nephrologist, or cardiologist. The intended interaction is collaborative. Rather than presenting a binary pop-up to accept or dismiss, the care team brings a clinician a concise patient context, explains why one intervention appears especially important, and asks whether the clinician agrees. The provider remains the decision-maker.

Toth-Manikowski said that distinction has changed the response from clinicians. A provider who cannot review an unranked inventory of gaps may still welcome a call identifying a rapidly progressing patient who has been lost to follow-up.

The broader evidence also counsels humility. In a 2024 cluster-randomized trial of 1,596 high-risk patients with CKD, an electronic health record–based, multidisciplinary population health intervention increased exposure to ACE inhibitors or angiotensin receptor blockers but did not reduce CKD progression over a median 17-month follow-up. The study’s mixed results are a reminder that identifying gaps and changing processes do not automatically produce better long-term outcomes. Programs need rigorous evaluation of recommendation quality, clinician uptake, intervention completion, equity, safety, and patient outcomes—not just model accuracy.

AI should surface signals, not substitute for judgment

Healthmap’s current next-best-action work, Vattamattam explained, largely uses machine learning and statistical methods to rank care opportunities from available data. Clinical experts and technical teams interrogate the output together. A plausible-looking answer is not accepted simply because a model produced it.

Generative AI may extend that system by extracting signals from less structured information. Vattamattam offered a hypothetical example: during a rapport-building call, a patient mentions falling twice while playing with grandchildren. A generative system could flag the comment and suggest that the nurse consider an appropriate blood-pressure or fall-risk protocol, potentially changing the priority established from structured data alone.

The same approach could surface nonclinical barriers such as transportation or behavioral-health needs. Its role would be to make relevant information harder to miss—not to diagnose, prescribe, or override the care professional.

“Always keeping a human in the loop” is essential

That safeguard matters because generative models can be inconsistent, and healthcare context is highly individual. If two people have the same CKD stage, they may still require different next actions because of their disease trajectory, comorbidities, medications, access barriers, preferences, or current relationships with clinicians. Personalization is not merely adding more variables to a prediction; it is applying clinical judgment to the circumstances behind them.

Why prioritized CKD care matters beyond the kidney

Toth-Manikowski emphasized that earlier, better-coordinated CKD care may influence far more than renal outcomes. Patients with CKD frequently have interconnected cardiovascular and metabolic risks. A missed nephrology appointment may coexist with uncontrolled hypertension, poorly managed diabetes, or a medication gap.

Getting the right patient to the right clinician can be the entry point for subsequent laboratory testing, medication management, and follow-up. In that sense, an appointment is not the final intervention. It restores a relationship through which a sequence of interventions can occur.

This whole-person view also changes the economic logic. Vattamattam and Toth-Manikowski argued that appropriate spending should move toward outpatient care and away from avoidable emergency-department visits, hospitalizations, and unplanned dialysis starts. Those are the outcomes Healthmap’s clients hire it to pursue, they said.

Importantly, these are company leaders describing their model and expectations in a podcast—not independent proof that every recommended action will generate a particular saving. Their more interesting strategic point is that a ranked intervention creates a unit of work that can be measured. An organization can estimate the value associated with closing specific opportunities, set targets for frontline teams, and compare expected with realized results.

That creates a potential chain of accountability:
patient-level priority → clinician decision → completed intervention → measured clinical and financial outcome

Risk-based contracting can strengthen that chain when the care-management partner and client share responsibility for outcomes. Vattamattam contrasted that alignment with purchasing software alone, where the buyer may carry most of the burden of determining whether the technology changed care. For health plans and provider organizations evaluating a kidney-care partner, the episode suggests several questions:

1. How is a recommendation ranked for an individual patient?
2. Which data sources inform it, and how are missing or conflicting data handled?
3. Where did clinicians participate in model design and validation?
4. How does the workflow preserve clinician autonomy and prevent alert fatigue?
5. What happens after a provider accepts a recommendation?
6. Are outcomes measured at the model, workflow, clinical, patient-experience, and financial levels?
7. How are incentives shared when results differ from projections?

The strategic takeaway: prediction needs an operating model

Healthcare organizations have invested heavily in identifying risk. The next competitive advantage may come from deciding what to do with it. A next-best-action strategy is not simply another score, dashboard, or alert. Done well, it is an operating model for converting a complex patient record into a prioritized conversation, preserving the clinician’s authority, and coordinating the work that follows. It requires technical capability, but also clinical governance, multidisciplinary collaboration, careful implementation, and incentives tied to outcomes.

The hardest part is not producing a longer list of what evidence-based care could include. It is helping a patient and clinician begin with the action that matters most—and then learning whether it worked.