Gemini flags 69% of held-out thermal-video falls

In an arXiv preprint submitted Oct. 21, Marcos Junca and Mauro Junca of Senecta said Gemini 3.8 Flash identified 69% of held-out subjects’ staged falls in low-resolution thermal video.

The Oct. 21 preprint frames multimodal LLMs, models that process video and other inputs, as a lower-training alternative to conventional detectors, which the authors said produced 14.2 and 23.9 false alarms per day in the home.

The authors tested a production multimodal model, Gemini 3.8 Flash, as the decision stage in an event-triggered fall screener. The setup used one request for each short thermal clip, adding the soundtrack when one existed, and compared the answers with Gemini 3.7 Flash on the same requests.

In a public 32x24 thermal dataset, the researchers manually located 438 staged falls and 563 non-fall windows. With bed falls excluded, the deployed request called 67% of falls across all subjects and none of 499 non-fall windows, including cases in which people lay on the floor and then stood up.

The home test was smaller and harder. Across 76 hours of ordinary activity in an occupied room recorded at 160x120 resolution with sound, Gemini 3.8 Flash called none of 395 ordinary-life clips a fall, which the authors translated to a 95% upper bound of 1.2 false alarms per day. It detected 4 of the 9 staged falls visible to the camera.

Prompt wording changed the result. The authors said a cautionary instruction — "a person who goes out of view ... is not a fall" — explained many misses.

Removing that sentence raised sensitivity, the share of true falls detected, from 69% to 83% on held-out subjects with Gemini 3.8 Flash and from 80% to 92% with Gemini 3.7 Flash.

The same prompt change lifted the home result from 4 detections to all 9 visible staged falls, while still producing no false fall calls during ordinary living, according to the preprint. The tradeoff was that deliberate lie-downs were then classified as falls.

The paper also highlights limits that matter for deployment. Sensitivity fell to 34% at 8x6 pixels, identical requests returned different answers on 4% of windows, and when the model did call a fall, its explanation placed the event within one second of the annotated landing 99% of the time.

The comparison with conventional detectors was mixed. Two detectors trained on the laboratory data caught nearly every held-out laboratory fall, the authors said, but generated far more home false alarms than the LLM screener: 14.2 and 23.9 per day.

The authors argue thermal cameras can monitor older people living alone without capturing faces, but the findings do not establish clinical readiness. The paper is an arXiv preprint and has not been peer reviewed, and the home evidence covers one occupied room and 9 visible staged falls.

The authors said they released frame-level fall annotations for the public dataset and every model answer on it. The preprint lists 13 pages, 6 figures and a data link hosted by Senecta.

Source: arXiv preprint, arXiv cs.AI, Oct. 22, 2026.