This was a contracting project I completed for a small medical technology company.

In Fall 2024, I met with the CEO and CTO of a San Diego HealthTech company that wanted to integrate machine learning into their product. Their product is a smart home solution (think tablets and connected sensors) for patients who require heavily involved daily caretaker support. Specifically, the company’s goal is to increase independence for patients while reducing staffing needs from the caretaker’s end.

The company had several sources of data they felt weren’t being fully taken advantage of — notably motion data from their sensors — but the scope wasn’t clear.

So after a few conversations to understand their product, customer, and data, we decided on a solution that would use ML to create high-level insights that could be quickly and easily interpreted by caretakers.

<aside> 💡

The product is currently being integrated into their platform and will be launching for beta customers in the future.

</aside>

What were the Project Priorities?

1. Quickly Interpretable: Results should be valuable while being quick and easy to understand
2. Generalizable: Must work across many patients — even patients with abnormal routines
3. Low-Touch: Requires very little work from the caretaker to set-up

Sample histogram for a random day. The morning and evening windows are estimated using kernel regression and various tunable processing functions.

What Does it Do?

The product uses bedroom motion sensor data (it is at caretaker discretion to choose to install motion sensors) to do the following:

• Estimate when morning and evening windows are
  • This is done using kernel regression with various data processing functions. Much of these functions were built around tunable hyperparameters, and tuning these allowed me to increase accuracy on hand-labeled test data from ~20% to ~90%. The final parameter set was further tested on held out data and generalized well
• Generate historical averages for your morning and evening windows
• Use these windows to compute various metrics of interest (e.g., nighttime activity) and compute several anomaly flags using this information (e.g., excessive nighttime activity, late evening, short sleep window, etc.). These flags are computed using a variety of statistical techniques
• Automatically generate charts with the activity histogram and labeled windows
• Finally, generate a routine score using this information to provide a simple metric for caretakers to understand patient behavior

## Sample Output of Anomaly Flags (Duration in Hours) 
Yesterday Morning Status: Typical
Yesterday Evening Status: Typical
Today Morning Status: Late
Today Evening Status: Typical
Night Events Count: 1
Night Events Duration: 0.099
Night Length: 10.597
Night Activity Status: Minimal
Night Length Status: Long
Average Night Length: 8.352