Baseline

Inspiration

In older adults, an infection often shows up first as confusion, not fever. Fever may be absent in 20 to 30 percent of older adults with a serious infection, according to a 1996 review in Infectious Disease Clinics of North America. When researchers studied 817 older adults who came to U.S. emergency departments with COVID-19, 28 percent were delirious on arrival. More than a third of those had none of the usual signs, such as fever or shortness of breath, the 2020 study in JAMA Network Open found.

That confusion (delirium) is easy to miss. In a 2001 study in Archives of Internal Medicine, nurses recognized delirium in only 31 percent of the hospitalized older patients who had it. It is even harder to spot in people with dementia. A 2008 study of long-term care residents with dementia found that 71.5 percent had delirium, yet nurses detected it in just 13 percent of cases.

Care aides often notice that someone "seems off" days before anything reaches the chart. In one nursing-home study published in the Journal of the American Geriatrics Society, aides documented signs of illness an average of five days before the chart did. But "seems off" is not something a system can track. On a night shift one nurse may cover a whole unit: Ontario's Auditor General found evening and overnight ratios as high as one nurse to 80 residents. And agency staff do not know what anyone's normal looks like. Their share of hours worked in Ontario long-term care nearly doubled in a single year, from 7.6 to 14.9 percent, according to Ministry of Health figures obtained by The Canadian Press.

PointClickCare's "Catch It Early" challenge asked for a way to catch a confused moment early and to explain why something was flagged. We built Baseline to make a change from someone's usual self visible, measurable, and easier to explain.

What it does

Baseline supports care teams looking after older adults, including people living with dementia. It measures changes in a resident's speech against their personal baseline: the range of speech patterns seen in their earlier checks. This keeps the focus on what has changed for that individual, even when memory or communication difficulties are already part of daily life. Care teams can review the change alongside relevant care notes and prepare a handoff.

How it works

A care aide opens a resident's check and records them counting and saying the days of the week during a 30-second recording window. Baseline checks the recording, turns speech into text, and measures patterns such as speaking speed and pauses. Earlier checks establish that resident's personal baseline. Each new check is measured against it to detect changes in how they speak.

The result reaches the nurse's overview, where residents needing attention are easier to find. Opening a resident shows the change over time alongside relevant notes and links to the original records. The nurse reviews that context and decides what to do next. The resident simply receives a thank-you, without a clinical score.

If a recording is unusable or there is too little history, Baseline says so. A change is a reason to look closer; it does not diagnose dementia, delirium, or their causes.

Challenges and what we learned

The hardest part was making a change meaningful. Everyone speaks differently, so a measurement needs the context of that person's usual patterns. We learned to show the personal baseline, the change from it, and the supporting notes together.

What we're proud of

We connected a simple voice check to personal-baseline comparisons and an evidence-backed review in one working prototype, demonstrated with fictional residents.

What's next

Test the experience with consenting volunteers and care professionals, including how to make voice checks comfortable and accessible for people living with dementia. Baseline is a decision-support prototype; it has not been clinically validated.

How we built it

A care aide records a 30-second voice check in the browser. Baseline checks the audio quality, transcribes the speech, and measures speech patterns. It compares these measurements with the resident’s personal baseline, established from earlier voice checks. Any changes are displayed alongside relevant care notes, so the nurse can review the supporting evidence and decide what to do next.

The app runs on Google Cloud. If the audio is unusable or there is too little history, it asks for another check or shows that more information is needed.

Built with Python, FastAPI, PostgreSQL, TimescaleDB, FFmpeg, Whisper, HTML, CSS, JavaScript, Docker, and Caddy.