When we started building the risk scoring engine at Patientrig, the first question we had to answer precisely was: which EHR data elements actually predict patient disengagement, as opposed to which ones tell you a patient has already disengaged? The difference sounds subtle, but it drives everything about how you design an alerting system. If your features are primarily diagnostic — they describe a care gap that already exists — your model tells care teams what they already know, just in a different format. If your features are genuinely leading indicators, you have a tool that changes the intervention window.
After about a year of working through this problem on real clinical data, here's what I can say with reasonable confidence: the predictive signals for disengagement are structurally different from the diagnostic signals most EHR-based population health tools surface. They're not in the lab values or the problem list. They're in the behavioral and transactional data — appointments, prescriptions, ADT events, and communication patterns.
The Diagnostic-Predictive Distinction in EHR Data
Most clinical risk scores used in population health — HCC scores, RAF (Risk Adjustment Factor) scores, and legacy utilization-based risk stratification — are built primarily from diagnostic and claims data: ICD codes, CPT procedure codes, prior hospitalization frequency, and comorbidity indices like the Charlson or Elixhauser. These are excellent for identifying which patients have the most complex clinical burden. They're poor predictors of which patients are about to disengage.
The reason is structural. A CHF patient with an HCC score in the top decile of your panel has always had that score — it tells you about their clinical complexity, not about whether something changed last week. The signal you need for timely disengagement prediction is a change in the patient's behavioral pattern relative to their own baseline, not a snapshot of their clinical acuity.
This is the core distinction we built around: features that capture deviation from a patient's individual engagement baseline, layered on top of (not instead of) static clinical risk factors.
The Data Elements That Actually Move the Needle
ADT Events: More Predictive Than Most Care Teams Realize
Admit-Discharge-Transfer feed data is commonly used for care transitions — triggering post-discharge follow-up workflows, generating care management alerts for hospitalizations. What gets used less frequently is ADT data as a continuous engagement monitoring signal. Specifically, the patterns that show up before disengagement in our data:
- ED visits without subsequent PCP follow-up within 14 days: This pattern — repeated ED utilization without reconnection to primary care — is a strong indicator that the patient is using the ED as their primary care touchpoint and has effectively disengaged from ambulatory management. It tends to precede formal HEDIS gap accumulation by two to four months.
- Inpatient discharge without pharmacy contact within 72 hours: When a patient is discharged with a new or modified medication regimen and no pharmacy claim appears within 72 hours, the discharge medication reconciliation is incomplete regardless of what the discharge paperwork says. This is one of the clearest leading indicators of the refill gap pattern that generates CHF readmissions.
- Observation status admissions: Observation encounters don't trigger the same post-discharge workflows as inpatient admissions in most EHR configurations. Patients who have multiple observation status visits for the same condition without a care management flag are frequently missed in post-acute monitoring workflows.
Rx Fill History: The Highest-Signal Feature We Work With
PBM claims data and integrated pharmacy feeds contain the most reliable leading indicators of disengagement we've found. The key metrics:
Days supply gap on maintenance medications. The number of days between when a patient's current supply of a chronic maintenance medication should run out and when the next fill claim appears is, in our experience, the single most predictive feature for short-term disengagement. A 10-day gap on metformin for a stable diabetic is different from a 10-day gap on lisinopril for a CHF patient post-discharge, but both warrant different levels of alert logic. We calculate this as a rolling metric against each patient's own historical refill cadence, not against the label-indicated days supply, because real refill behavior often doesn't match label timing.
New prescription fill failure. When a new prescription is written at discharge or after a medication change, a fill claim that doesn't appear within five to seven days is highly predictive of non-initiation. Prior authorization delays are a common cause; so is patient cost-sharing concern (particularly for specialty medications like GLP-1 agonists or newer SGLT2 inhibitors). The EHR often shows the prescription as "active" with no indication that it was never filled. The only way to detect this is by matching the prescribing event in the EHR against PBM claims.
Medication switches and supply fragmentation. When a patient is switched from one drug to another within a class (e.g., from enalapril to lisinopril following a formulary change), there's often a transient period of supply confusion where they have days of one medication left, haven't yet filled the new one, and may be partially adherent, fully non-adherent, or inadvertently doubling. This shows up as an anomaly in the days supply calculation that standard PDC calculations often smooth over.
Appointment Sequence Patterns
Single no-shows are weak predictors of disengagement — they're too common and too confounded by scheduling system artifacts (automatic cancellations, same-day reschedules) to be reliable signals on their own. The appointment sequence features that have predictive value:
- Two consecutive no-shows on the same appointment type within 90 days, with no rescheduling for either
- No-show followed by no response to outreach within 48 hours (this combination is much stronger than either signal alone)
- Increasing interval between kept appointments — a patient who was seen every 90 days and is now 160 days since last contact, with the next appointment still two months out
- Specialist referral without a corresponding claim within 60 days — indicating either that the referral wasn't followed through, or that the patient attended a different provider outside the network (which has its own data visibility implications)
What Doesn't Predict Disengagement (Counter-intuitive Findings)
Lab values — specifically abnormal lab values — are surprisingly weak features for predicting disengagement, at least as point-in-time measurements. An HbA1c of 9.5% tells you a diabetic patient has poor glycemic control; it doesn't tell you whether they are currently engaged with their care team or not. We've seen patients with very high HbA1c values who are actively engaged, showing up regularly, and working with their team on glycemic management — and patients with moderately elevated HbA1c who have completely disengaged. The lab value describes the outcome of adherence behavior; it doesn't predict the behavior going forward.
Similarly, Charlson comorbidity index and HCC score are weak features for near-term disengagement prediction. High clinical complexity and high disengagement risk are correlated at the population level, but at the individual patient level — which is where care management operates — the correlation is loose enough that using clinical complexity as a proxy for disengagement risk leads to systematically over-triaging the most complex patients and undertriaging the patients who are actively losing engagement on a simpler condition profile.
We're not saying clinical complexity scoring is useless — it's valuable for resource allocation and care management intensity decisions. What we're saying is that it shouldn't be the primary input for real-time disengagement alerting.
Data Quality Issues That Affect Predictive Accuracy
Ten years of healthcare data engineering has taught me that the quality problems that matter most for disengagement prediction aren't the ones most data teams focus on. The common focus is on data completeness — ensuring all encounters are captured, all diagnosis codes are present, all lab results are reconciled. That matters for clinical decision support. For disengagement prediction, the relevant data quality issues are different:
PBM claims lag. PBM claims typically arrive 7 to 14 days after the fill event. For a post-discharge patient where the intervention window is 72 hours, this lag can make pharmacy-based alerting effectively useless unless you have a real-time pharmacy data integration (through an in-house dispensing feed, a contracted PBM integration, or an HIE that provides real-time Rx data). We spend significant time in implementations evaluating whether a health system's PBM data source is actually real-time or near-real-time, because the lag profile completely changes the alerting logic.
Patient demographic data accuracy. Phone number and address currency is a recurring problem that care teams underestimate. A phone number that's 18 months old for a patient who moved or changed carriers generates a false "outreach attempted, no response" event that looks like patient non-responsiveness but is actually a data quality failure. For disengagement models, systematically stale contact information creates a confounding pattern that needs to be accounted for explicitly.
ADT feed completeness across care settings. Most health systems have good ADT coverage for their own facilities. Coverage of ED visits and admissions at competing regional hospitals, urgent care centers, and SNFs is often incomplete or delayed. A patient who was seen at a competing system's ED and discharged two weeks ago may appear in your population health data as "last contact 4 months ago" — when actually they had a recent clinical encounter, just not in your network. This matters for disengagement scoring because a patient who appears disengaged based on your ADT feed may actually be engaged with a competing care system, which is a different care management problem than genuine disengagement.
Building Toward a Reliable Signal
The architecture that works for disengagement prediction combines a relatively small set of high-signal behavioral features — pharmacy fill patterns, ADT event sequences, appointment adherence trends, and outreach responsiveness history — with a patient-specific baseline that captures what's normal for that individual. The goal isn't to predict which patients are clinically complex; it's to predict which patients are currently departing from their own engagement pattern in a way that historically precedes a care gap.
For health systems evaluating tools for this use case, the questions that matter are: What is the primary data source for the risk score — diagnostic claims or behavioral transactional data? What is the latency profile of the underlying data feeds? Does the model score against a static population-level risk threshold or against each patient's individual engagement baseline?
These aren't abstract architectural questions. They determine whether the alerts the tool generates are actionable within the clinical intervention window — or whether they're confirming what the care team already knew, too late to change the outcome.
