New Model Performs Well Predicting Parkinson’s Sentinel Falls

Falls are common for people with Parkinson’s disease. Even first falls can lead to injury, hospital stays, loss of independence and reduced quality of life. They also can be difficult to predict. Traditional prediction models are best at identifying risk only after an individual has already fallen.

Cleveland Clinic researchers have been aiming to change that by developing a tool that efficiently predicts the first fall using routinely collected data. Their recent study in Movement Disorders Clinical Practice, Closing the Gate Before the Horse is out of the Barn, describes results of a novel prediction model that incorporates scores from standard cognitive tests, which have not been used in traditional tools.

“Most models do not do a very good job of predicting the first fall in patients with Parkinson's disease or, frankly, older adults,” says Jay Alberts, PhD, Vice Chair of Innovation in the Neurological Institute. “Most rely on previous fall history. So now we have come up with a model that can predict first falls in patients with Parkinson’s disease, and that is important so we can potentially send those patients to physical therapy or just let them know.”

The research also highlights outcomes of participants’ first falls and underscores that falling shouldn’t be considered inevitable.

“It’s problematic to assume we can pay attention only after someone has already fallen,” Dr. Alberts says. “We showed that the first fall is consequential.”

Waiting Room of the Future

In the past, fall prediction models have relied heavily on motor-related data that often requires instrumented or recall-based assessments. The researchers first sought to improve on that by developing a cognitive-based model using data from the Cyclical Lower Extremity Exercise for PD II (CYCLE-II) trial. This CYCLE-II model of fall prediction identified slower cognitive processing speed and visual memory and longer disease duration as key risk indicators.

Researchers built on that to develop a reliable first-fall predictive model with data routinely collected and uploaded to the electronic medical record (EMR) via Cleveland Clinic’s Waiting Room of the Future (WROTF).

The WROTF collects data from patients using iPads under supervision to complete test modules while they wait for routine neurological appointments. The modules assess cognitive processing speed, visual memory, manual dexterity and 25-foot comfortable walking speed.

Researchers applied least absolute shrinkage and selection operator (LASSO) regression to WROTF-collected information and variables included in the EMR. They knew that if this new tool – the LASSO WROTF model – was to be effective, its strengths would be efficient use of routinely collected data and the ability to seamlessly feed results back into the EMR.

The researchers also sought to:

• externally validate CYCLE-II Cognitive Model using WROTF data
• compare the performances of LASSO WROTF, CYCLE-II, and a three-factor motor-based prediction model developed by Paul, et al, in 2013 (the Paul model).

Full cohort analysis

Of 997 study participants, 570 had had no falls in the previous year, 367 had fallen, and 60 had an unknown fall. Those who had fallen tended to have had Parkinson’s for a longer time and had higher scores on the Movement Disorders Society Unified Parkinson Disease Rating Scale (MDS‐UPDRS) Part III compared to those who were fall-naïve. They also performed worse on cognitive processing and visual memory tests than non-fallers.

Future fallers followed similar patterns. They also took higher doses of levodopa and had slower walking speeds compared to those who did not fall.

The WROTF LASSO model showed the best overall discrimination of the models evaluated, with a cross-validated AUC of 0.81, indicating better predictive performance than the motor-only model. It identified history of previous falls as by far the strongest predictor of future falls, followed by slower processing speed, slower gait speed, longer disease duration, and, to a lesser extent, greater motor severity and higher dopaminergic medication burden.

The original CYCLE-II cognitive model retained moderate predictive ability when it was applied to the WROTF cohort (AUC of 0.69). As previously noted, when the investigators refit the model for the WROTF cohort using the same predictors, they found that slower cognitive processing speed, worse visual memory and longer disease duration were associated with greater fall risk. Model performance was essentially unchanged after refitting, with a cross-validated AUC of 0.70.

The Paul motor model performed well overall (AUC 0.77) but was driven primarily by prior fall history, followed by gait freezing and slower gait speed.

Fall-naïve patients

Overall, model performance in fall-naïve patients was attenuated compared to the full cohort, reflecting the absence of fall history as a dominant predictor. The cognitive model (CYCLE-II) demonstrated relatively stable discrimination, while the motor-based Paul model showed a more pronounced decline in performance. The WROTF LASSO model maintained the highest and most balanced performance (AUC = 0.70), suggesting improved generalizability in identifying first-fall risk.

At the time of WROTF data collection, 203 participants were fall-naïve and went on to fall at some point during the next year. Nearly half required emergency department care, while about 12% were hospitalized, 3% went to skilled nursing and 7% needed new assistive devices.

“There are costs associated with a first fall and there are changes in quality of life associated with that first fall, and even in terms of living situation for a small percentage of the population,” says Dr. Alberts.

Clinical integration

Because a first fall often marks a turning point in a patient's independence and quality of life, having objective data to prompt earlier intervention provides an invaluable opportunity to activate patient-specific preventive measures, says Saar Anis, MD, a clinical Fellow and first author on the research.

“When our assessment identifies a patient at elevated risk for a first fall, it opens a proactive conversation about prevention before that sentinel event occurs,” says Dr. Anis. “I can tailor my approach based on the specific risk factors, whether that means referring to physical therapy for gait and balance training, adjusting medications that may affect mobility or cognition, addressing orthostatic hypotension, or connecting patients with occupational therapy for home safety evaluations. The key is that we're intervening preventively rather than reactively.”

The most important takeaway from the research is that patients at elevated risk for a first fall can be immediately identified using assessments already integrated into routine clinical care.

“Because previous fall prediction models relied heavily on prior-fall history, we were always playing catch-up. Our approach combines cognitive and motor measures collected in the waiting room before appointments, and this data flows directly into the electronic health record,” says Dr. Anis. “This means fall-risk assessment becomes part of routine care rather than an additional burden.”

And, he adds, it is notable that cognitive measures, especially processing speed, emerged as key predictors alongside traditional motor assessments. “This reinforces that falls in Parkinson's disease result from complex interactions between motor, cognitive and environmental factors.”