Predictive Fall Risk Stratification Tools for Nursing Home Residents — Vendor Landscape

Nursing home fall risk stratification now has two generations of tools, static scoring instruments and dynamic, EHR-driven machine learning tools. For facilities evaluating vendors, the essential distinction involves predictive accuracy and whether a tool can detect changing resident risk from data already being collected, fit nursing workflows, and support continuous oversight without adding unsustainable documentation work.

Why falls are a different problem in nursing homes than in acute care

Nursing home residents combine high dependency, frailty, and multiple chronic conditions in ways that complicate fall risk assessment. Falls can cause severe physical injuries, speed up functional decline, and leave residents needing more care. Those consequences compound across a long-stay population. Yet research specific to nursing home fall risk assessment remains scarce compared with the evidence available in acute care, limiting how confidently facilities can apply findings from other settings.

The instruments themselves reflect that imbalance in development. STRATIFY and Hendrich II were designed and tested in acute care or hospital subacute settings, while the Morse Fall Scale was developed across acute, rehabilitation, and nursing home settings. Their origins make setting-specific performance important for their use in LTC. Limited staffing and scarce resources impose a further constraint, because a nursing home needs an assessment process that works within its care delivery capacity.

What static scoring tools actually measure — and where they break down in LTC

The Morse Fall Scale (MFS), STRATIFY, Hendrich II, and Timed Up and Go Test (TUGT) assess a fixed set of factors at a particular time. The score stays the same until someone repeats the assessment. That creates a gap between measurement and current condition. For residents whose condition shifts daily, the practical limit is not only which factors an instrument includes, but how quickly it can reflect a change.

A 2025 prospective cohort study across four Chinese nursing homes demonstrated a second limitation. At their original cutoffs, MFS and TUGT showed high sensitivity but very low specificity, identifying many potential fallers while also flagging many residents who would not fall. STRATIFY and Hendrich II showed the reverse pattern. Their high specificity came with very low sensitivity, leaving many eventual fallers unidentified. Adjusting cutoffs improved the balance, and TUGT achieved the strongest overall predictive accuracy, but meaningful gaps remained.

The comparison with nursing judgment is particularly important. A cluster-randomised controlled trial found unsatisfactory predictive values for both assessment tools and nurses’ unassisted clinical judgment, with no tool superior to judgment alone. A separate retrospective study in geriatric evaluation and management units found that tools worked better only with cutpoints set for each institution, work that many nursing homes cannot easily keep up. Static instruments still have a place in formal assessment. But the work of documenting their scores needs to be justified when their risk categories may predict little more than an experienced nurse’s assessment.

How ML-based tools work and why EHR data changes what is predictable

Machine learning changes the assessment mechanism by using a larger set of EHR variables and updating risk as new documentation becomes available. Instead of waiting for a repeated checklist, a model can incorporate changes in routine clinical data. Vital signs are a particularly useful signal. Models incorporating them have shown better fall risk surveillance than models without them, while conventional static instruments do not capture this source of changing risk.

Extreme Gradient Boosting, or XGBoost, can use many EHR features to make short-term predictions and has shown better performance than conventional fall assessments. The Juniper Communities study provides a concrete example of the broader approach. It tested three ML models intended to predict 3-month fall risk using a proprietary EHR database spanning 2007–2021 and thousands of individuals. The study included people in skilled nursing, independent living, and assisted living. Its findings support dynamic, individualized care practices, although that mixed-setting evidence is not definitive validation across nursing homes.

The model at the TigerPlace senior living facility also used nursing note text and medication records as data inputs. Its deep-learning approach achieved improvements in accuracy, sensitivity, and specificity over traditional tools. These findings offer preliminary evidence of promise at scale. Operationally, the distinction is that updated scores can become a byproduct of existing charting, rather than another assessment nurses must administer.

The frontier of unstructured note analysis and what it adds to structured EHR data

Most ML fall prediction models in use today draw mainly on structured EHR variables. Those fields capture what was measured or coded, but not everything a nurse observed. Notes contain a different kind of clinical signal. Behavioral cues, gait changes, and agitation may be recorded in narrative documentation but left out of a structured field, so structured-only models cannot use them.

Natural language processing, or NLP, makes some of that narrative content available for prediction. A combined model using structured clinical variables and NLP-extracted topics from nursing notes outperformed either component alone. The finding supports an additive role for notes. Notes supply observations omitted from vital signs, medications, and other coded information.

A Dutch hospital study also demonstrated the value of narrative records for identifying falls that had already occurred. Its algorithm, based on RobBERT, a fine-tuned Dutch BERT model, identified a substantially higher proportion of fall incidents than incident reports alone while reducing review time by more than ninety percent.

The regulatory shift that turned fall prediction from a quality initiative into a compliance imperative

CMS revised the Special Focus Facility (SFF) selection criteria in January 2026 to include falls data as a direct selection factor. When State Agencies choose between two facilities with similar compliance histories, they are directed to select the one with higher fall rates. Falls therefore have a named consequence for regulatory scrutiny. The change raises the operational stakes of detecting risk and sustaining prevention, rather than treating assessment as an isolated quality exercise.

CMS is starting a nationwide Risk-Based Survey (RBS) process for nursing homes that qualify, beginning September 8, 2026. The process is designed to let State Survey Agencies redirect resources toward facilities where residents face greater risk, while only a small minority of nursing homes currently qualify for the favorable tier. This makes sustained performance increasingly consequential between surveys. It strengthens the case for reliable, ongoing risk oversight.

OIG supplied important context in September 2025, reporting that nursing homes failed to report nearly half of falls with major injury and hospitalization among residents enrolled in a federal health insurance program. That finding indicates that self-reported falls data understates actual incidence at the population level, while leaving reporting accuracy at any particular facility undetermined. The distinction is essential for fair interpretation. For operators, the resulting challenge extends beyond producing a risk score to demonstrating an ability to detect risk, intervene consistently, and learn from events. Dynamic monitoring can support that work, while an episodic assessment alone cannot demonstrate continuous performance.

What operator-led deployments reveal about real-world performance and its evidentiary limits

Signature HealthCARE, based in Louisville, KY, initially deployed an ambient ML monitoring system in four facilities and reported an eighty percent decrease in falls in buildings using fall prevention AI. The system operates as a virtual sitter program. Machine learning detects resident movement and notifies staff, while sitters can monitor multiple residents and initiate two-way communication with residents or staff after movement alerts. Signature later added the system to more buildings, showing confidence in its use beyond the first rollout.

Trilogy Health Services, also based in Louisville, KY, runs 145+ senior living communities and chose to build its own AI infrastructure. It made that choice to keep ownership of resident data and avoid licensing it to a third-party vendor. For high-risk residents, a 24-hour chest-worn monitoring device supports anticipatory care and has been associated with a substantial hospitalization reduction tracked weekly across a large cohort. Trilogy’s internal auditing tools, chatbots, and documentation search tools also show how it is bringing clinical AI together around resident data it owns.

Both cases are instructive, but their outcomes are operator self-reports presented at industry conferences. Neither establishes what another facility should expect from a purchased product. There is still too little evidence that these tools work in real-world LTC. During procurement, ask for the study design behind the result, treating rollout experience as evidence of operational feasibility without confusing it with controlled evidence of effectiveness.

How the two categories compare operationally — data requirements, workflow fit, and staffing reality

Static instruments place the update burden on staff. They are administered at admission and on scheduled reassessment cycles, and a new score requires another administration of the instrument. Their currency depends on that assessment schedule. Because the documentation is additional to routine charting, the operational question is whether the information gained justifies the repeated work.

Structured-data ML models shift that burden toward data quality. They ingest coded EHR fields, including vital signs, medications, diagnoses, and ADL scores, and can update risk automatically as those fields change. They need consistent, complete structured charting to work. Where the required data is already recorded reliably, scoring does not require additional nursing documentation, but automation cannot substitute for the inputs the model needs.

Note-aware ML adds narrative extraction to that foundation. It can capture behavioral and observational signals without requiring nurses to write a separate assessment, because the source is their existing notes. The technical demands are greater, however. This subtype offers broader predictive potential than structured-only models, but its fit depends on the available narrative data and the facility’s capacity to implement it.

The KLAS Arch Collaborative’s 2025 report found that a large majority of responding acute care nurses lost time to duplicative or unhelpful documentation. It reinforces the workflow issue facing documentation-heavy care settings, where another data-entry requirement can undermine an otherwise attractive tool. Models that use existing charting spare staff extra documentation for scoring, but putting them into use still takes work.

Ambient monitoring, as used by Signature, gets its data from a different source. Sensor-generated signals require physical hardware, ongoing device management, and integration infrastructure. An evaluator must consider predictive performance, what data the tool needs, whether the facility generates it reliably, and whether adoption adds work to the nursing manager’s day or reduces it.

What a facility should establish before choosing a fall risk stratification tool

Before evaluating vendors, a facility should establish its requirements across four questions, beginning with how often risk must be updated. Does the proposed tool produce a static score or a continuously updated signal. A scheduled reassessment alone cannot keep a facility ready at all times when a resident’s condition changes between assessments. How often a tool updates is therefore a feature that must meet the needs of daily operations. It should be considered against both daily care needs and the demand for sustained oversight between surveys.

The second question concerns the data the tool consumes, including whether it reads structured fields alone or nursing notes as well. Facilities with detailed narrative documentation have an additional source of predictive information that structured-only tools cannot use. Facilities must also assess the work required to supply those inputs. A sustainable implementation should use existing charting rather than introduce new documentation workflows for scoring. Together, these questions establish whether the proposed intelligence is accessible from the facility’s actual records.

Facilities should distinguish operator-reported outcomes, internal pilots, and peer-reviewed studies conducted in LTC settings, giving weight to the study design behind headline reductions. Pilot periods should include defined outcome metrics before enterprise commitments. This is especially important where results from another care setting or an operator’s proprietary infrastructure are being used to support a broader sales claim.

Large operators are building proprietary data systems and ambient sensor networks, but mid-size and independent facilities may not have the money for those approaches. EHR-integrated, note-aware tools offer a practical direction by extracting more useful signals from documentation already being produced. Selection still depends on whether a vendor can turn that direction into a workable, locally evaluated system under the facility’s staffing and data constraints.