AI Fall Risk Prediction in Hospitals

The Morse Fall Scale, published by Janice Morse in 1989, scores six items: history of falling, secondary diagnosis, ambulatory aid, IV or saline lock, gait, and mental status. Totals range from 0 to 125, and many hospitals treat scores around 45 or higher as high risk, though thresholds vary. Other tools include the Hendrich II Fall Risk Model and the Johns Hopkins Fall Risk Assessment Tool. These scales are quick, but nurses usually score them only once a shift, they miss changes between assessments, and they label so many inpatients as at risk that the label loses meaning. Predictive models recalculate risk from EHR data as it changes. Inputs can include sedatives, opioids, diuretics and blood pressure medications, sodium levels, orthostatic vital signs, delirium screens, toileting needs, recent procedures and time of day. EHR vendors, including Epic, offer fall risk models, and some hospitals build their own. Video is a separate layer. Continuous virtual observation, offered by companies such as AvaSure, lets trained staff watch many rooms at once and speak to patients through a speaker. Newer systems add computer vision that spots posture changes such as sitting up or moving toward the bed edge. That gives earlier warning than pressure-pad bed alarms, which trigger only when weight comes off the mattress. A key misconception is that more alerts mean fewer falls. A large randomized trial of increased bed alarm use did not find a reduction in falls. A risk score says who might fall, not what to do. Falls drop when an alert leads to action: scheduled toileting, low beds, non-slip footwear, gait belts, physical therapy, medication review and staying with patients in the bathroom. Since 2008, CMS has not paid hospitals extra for falls with injury that happen during a stay, which gives hospitals a financial reason to prevent them.