Predictive Fall Risk Stratification Tools for Skilled Nursing Facilities

Predictive fall risk stratification tools read clinical data continuously, not at fixed intervals, so they can flag which residents are most likely to fall before an incident happens. Understanding how they work, what signals they track, and how they slot into daily nursing routines matters now because falls data has become a direct regulatory trigger for skilled nursing facilities in 2026.

Why falls became a direct regulatory trigger in 2026

A facility's reported fall profile now shapes its regulatory exposure in a way it did not before this year. In January 2026, CMS revised the selection criteria for the Special Focus Facility program, so that falls data is now weighed alongside survey scores when state agencies decide which facilities to target for heightened scrutiny. That replaced the prior arrangement, in place since 2022, where staffing levels data served as the tiebreaker when facilities had similar compliance histories. Under the new rule, where two facilities look similar on paper, the agency is directed to select the one with the higher fall rate, turning a single metric into a determinant of which nursing homes get flagged for closer oversight.

CMS also widened what counts as a reportable fall. Updated MDS reporting requirements now capture intercepted falls and falls caused by external forces, categories that previously went unrecorded in many facilities. A facility can see its reported fall numbers climb even though nothing has changed in how often residents are actually falling, simply because the counting rules caught up to clinical reality. That creates two separate but linked obligations running through the same survey cycle: accurate fall reporting now affects SFF exposure, and MDS compliance more broadly affects eligibility for adjustments tied to a separate value-based scoring program. The same documentation failure, in other words, can cost a facility on both fronts at once. None of this works, though, unless the underlying fall data is trustworthy, and that is where the regulatory logic runs into a problem CMS did not create but is now leaning on anyway.

The documentation gap that undermines existing fall data

The fall data feeding this new regulatory apparatus is, at many facilities, badly incomplete. A federal oversight review found that close to half of serious falls, meaning those that led to hospitalization, never made it into MDS records during the period examined. The same review found something harder to explain away: facilities reporting the lowest fall rates were also the least likely to have actually reported the falls investigators checked against their records.

That finding flips the logic regulators are now building policy around. A clean-looking fall record is more likely to signal gaps in reporting than genuinely low incidence, so the SFF tiebreaker CMS introduced could end up rewarding the facilities with the weakest documentation discipline. Two pressures now converge on the same survey cycle as a result. Facilities have to lower their actual fall rates, tighten the accuracy of how falls get logged, and document prevention efforts in enough detail to withstand scrutiny, all at once, with no separate runway to fix one before the other becomes urgent. That problem starts with the tools most facilities use to assess fall risk, which were never built to produce the kind of continuous, trustworthy signal this new regulatory environment demands.

Why periodic, structured-only assessment tools leave a predictive gap

The fall risk tools in common use across skilled nursing, the Morse Fall Scale, STRATIFY, and the Hendrich II Fall Risk Model among them, share a design limitation rooted in how and when they collect information. Each depends on a clinician's subjective evaluation at a fixed point in time, scored against static parameters that do not update between assessments. That structure produces a known failure mode: a large share of residents end up classified as high-risk under scales like the Morse Fall Scale, a uniform label that gives nursing staff little basis for deciding which of those residents needs which specific intervention.

The deeper issue is temporal. A resident's gait can deteriorate, hydration can drop, or a new medication can start affecting balance in the days between one scheduled assessment and the next, and none of that appears in the record until the next scored evaluation happens or until a fall forces the issue. Skilled nursing facilities feel this gap more acutely than hospitals do, because staffing levels, high resident acuity, and turnover make it hard to sustain repeated manual assessment cycles across every shift. A set of falls prevention guidelines issued in 2022 by a global multidisciplinary initiative already pointed toward this limitation: the guidelines recommended multifactorial fall risk assessment and advised against relying on scored screening tools alone in hospital settings. None of this makes the existing scales useless. It does mean they capture a snapshot where the clinical problem is continuous, which is precisely the gap that predictive models are built to close.

Using EHR data for continuous risk monitoring

Predictive fall risk models read data that facilities already generate routinely through the electronic health record, continuously, so a change in a resident's condition appears in the record as it happens. That continuous read is the central design difference separating these tools from the Morse Fall Scale or Hendrich II. Research using EHR data for short-term fall prediction suggests these models can support dynamic care practices that respond to changes in an individual resident's risk profile within senior care settings, a kind of responsiveness a scheduled tool structurally cannot offer no matter how often it gets re-administered.

One architecture built specifically for the nursing home setting illustrates how this works without requiring new infrastructure. One hybrid predictive model, trained on the Long-Term Care Minimum Dataset, has been shown to outperform 22 fall risk assessment tools previously evaluated in nursing homes, with its advantage most pronounced for shorter prediction windows. It uses data fields already present in the MDS, the same dataset nearly every U.S. skilled nursing facility already files, so you don't need to build new data collection pipelines or ask staff to fill out additional forms. Other emerging architectures go further: they process vital sign trends through temporal analysis, nursing notes through natural language processing, and clinical context through parallel channels, and these combine into a risk score paired with an explanation of what drove it. The explanation matters as much as the score. A number alone tells a nurse that a resident is at elevated risk; an explanation tells her what to act on, which is the difference between a model that gets used and one that gets ignored. The strongest of those parallel channels, and the one most often left out of fall prediction entirely, is the nursing note itself.

The fall risk signal in unstructured nursing notes

Nursing notes contain observations about gait, behavior, pain, and confusion that never make it into a structured assessment field, and that omission is a central reason conventional tools miss so many falls before they happen. Published research has found that nursing notes in the EHR hold fall risk information, and facilities still underuse it as a predictive data source. The information is already in the chart. The barrier is that it sits in free text that a structured-field tool has no way to read.

Consider the clinical logic at work. A nurse who writes that a resident "seemed unsteady getting up from the chair" or "appeared confused at medication time" has just recorded a real, current change in that resident's fall risk. A structured mobility field scored at the last MDS cycle captures none of that nuance, because it was never designed to. The input already exists in the notes nurses write as part of routine care, so no new form, workflow, or end-of-shift field has to be added to generate it, which makes natural language processing the strongest practical argument for adoption in this setting. That said, the quality of the signal depends on the quality of the writing. NLP accuracy on clinical notes depends on consistency, and charting habits vary by facility and by shift, so a model trained on clean, detailed documentation can perform worse where notes run short or abbreviated. That is a real implementation constraint rather than a reason to dismiss the approach, and it points directly to what these models are actually scanning for once they can read the notes.

What signals predictive models prioritize

Predictive models give weight to signals that shift over days or hours, not to fixed traits that stay the same between assessments. Clinical research on fall prediction tools points to a consistent set of dynamic indicators: a recent fall or near-fall, a change in gait or mobility, a new psychotropic or sedating medication, altered mental status, or a change in toileting frequency or urgency. Each of these can appear in a nursing note well before it ever reaches a structured field.

Medication changes stand out as a particularly strong signal. Polypharmacy and sedating medications are established, well-documented fall risk factors, but a new prescription or dose adjustment is a discrete, dated event. A continuously running model can register that change in the medication record the same day it occurs, where a periodic tool simply waits for its next scheduled pass. Narrative research on fall risk assessment in nursing homes reinforces a related point: effective prevention depends on identifying which specific factors are driving risk for a given resident, the individualized picture, rather than relying on a single composite score that treats every high-risk resident the same way. Behavioral shifts carry similar weight. Increased agitation, new resistance to care, or nighttime wandering documented in a note can flag risk days ahead of an actual fall, so staff get a window to intervene that a scheduled assessment simply does not provide. Continuous AI monitoring is beginning to add another kind of signal entirely: research using AI-derived bed and chair exposure time has estimated position-specific fall rates normalized for how long a resident actually spends in a given position, producing an incident measure independent of staff-reported logs. That measure matters directly for the documentation gap described earlier, since it offers a check on fall counts that does not depend on a nurse remembering to file a report.

The sensitivity-specificity tradeoff and the real limits of AI fall prediction

Better overall discrimination does not mean a model catches every faller, and that tradeoff has direct consequences for how a facility runs its floor. A 2025 validation study put the tradeoff in concrete terms: the best-performing AI model in the study beat a four-item nursing assessment tool on overall discrimination and specificity, but the simpler nursing tool retained higher sensitivity. The AI model generated fewer false alarms, but it also missed more residents who went on to fall than the simpler tool did.

That tradeoff raises an operational cost on an understaffed floor. A model tuned for high sensitivity, built to catch as many potential fallers as possible, generates more alerts for staff to evaluate, and on a unit where nurses already carry a heavy documentation load, a high rate of false positives can push a team toward ignoring the alerts altogether, erasing whatever clinical benefit the model was supposed to deliver. The 2022 global multidisciplinary consensus on falls guidelines addressed a version of this problem from the standpoint of clinical governance: it recommended multifactorial fall risk assessment over purely numerical screening tools and insisted that any assessment account for local context and local resources. Threshold calibration, in other words, is not a setting a vendor gets to fix once and ship everywhere. Pairing the model with clinical judgment, rather than replacing one with the other, is what makes this work in practice. The Morse Fall Scale is strong at detailed stratification. The Clinical Judgment Method is strong at rapid screening. A predictive model is strong at continuous monitoring of change over time. Run together, the three cover different parts of the same risk picture instead of competing to replace each other, and that combination earns the technology a place on the floor instead of becoming one more alert nurses learn to swipe away.

Fitting continuous fall risk monitoring into SNF nursing workflows

A predictive model only holds value on a nursing floor if it runs on data staff are already producing, the charting, medication administration records, and vitals they already file as part of routine care. None of the mechanism described in this piece requires a new form at shift change or an extra field in the MDS. The model reads what already exists: nursing notes written in the normal course of care, medication orders entered as they happen, vital signs captured on the usual schedule.

What changes for a charge nurse is not the work of documentation but the output on the other end of it. Where a scheduled Morse Fall Scale assessment produces a single number every so many days, a continuously running model produces an updated risk picture as new information enters the record, paired with an explanation of which specific signal moved the needle: a new sedating medication, a note describing unsteadiness, a shift in toileting pattern. That explanation is what lets a nurse decide quickly whether the alert matches what she is already seeing in the resident, rather than forcing her to chase down a score with no clinical story behind it. Fitting into that rhythm, rather than disrupting it, separates a tool a facility actually uses from one that sits unopened in a dashboard: fall risk stratification that works continuously, reads what nurses are already writing, and gives care teams a reason to trust what it tells them.