Machine Learning · Healthcare
Wearable Sensors + Machine Learning: How a Single IMU Can Assess Mobility in Older Adults
By Iman Hosseini · 6 min read
Mobility decline is one of the strongest predictors of loss of independence in older adults — yet assessing it usually requires a clinic visit, specialised equipment and trained staff. Over the past few years, my research has focused on a simpler question: how much of that assessment can a single, inexpensive wearable sensor do?
Why a single IMU matters
An inertial measurement unit (IMU) combines an accelerometer and gyroscope in a package smaller than a coin. Most research systems use several of them, strapped across the body — accurate, but impractical outside the lab. If a single sensor can capture enough signal, screening becomes something you can do in a GP clinic, a pharmacy, or at home.
From raw signal to clinical insight
The pipeline looks deceptively simple: record a standard clinical movement like the five-time sit-to-stand test, extract features that describe how the movement was performed — its speed, smoothness, symmetry and variability — and train models to map those features onto clinical assessments of locomotive syndrome or fall risk.
The hard part is everything around the model. Sensor data from real older adults is messy: pauses mid-test, sensors shifting, movements that don't match the protocol. Robust preprocessing and honest validation matter far more than exotic architectures.
What we found
Across our published studies, single-IMU models achieved assessments comparable to standard clinical scoring — suggesting a future where mobility screening is cheap, objective and available anywhere. The models aren't replacing clinicians; they're extending their reach.
Engineering · Data Pipelines
From Messy Sensor Data to Reproducible Models: A Practical Pipeline
By Iman Hosseini · 5 min read
Every ML tutorial starts with a clean CSV. Every real project starts with a folder of mislabelled recordings, missing timestamps and three different file formats. Here's the pipeline structure that has saved me the most time across four research projects.
1. Ingest once, validate immediately
Raw data goes into an immutable "landing zone" and never gets edited by hand. An automated validation step checks sampling rates, durations and label consistency the moment data arrives — catching collection errors while you can still re-record, not three months later.
2. Make preprocessing a versioned artifact
Filtering, segmentation and feature extraction live in version-controlled code, not notebook cells. Each processed dataset records exactly which code version produced it. When a reviewer asks "what changed between these results?", the answer is a git diff, not archaeology.
3. Separate exploration from evaluation
Notebooks are for exploring; scripts are for results. Any number that might end up in a paper or a stakeholder report comes from a scripted, seeded, reproducible run. This one rule has prevented more embarrassment than any model improvement.
The payoff
These workflows cut manual processing time dramatically — but the real value is trust. When your pipeline is reproducible, every result you present is defensible. That's what makes models production-ready rather than just technically impressive.
Teaching · Career
What Six Years of Teaching Taught Me About Being a Better ML Engineer
By Iman Hosseini · 4 min read
I've spent six years tutoring signal processing, electronics and systems engineering at ANU while working as a data scientist. People sometimes treat teaching as a detour from "real" engineering work. I think it's the opposite — teaching is the best engineering training I've had.
Explaining forces understanding
You can use a Kalman filter without understanding it. You cannot teach one that way. Standing in front of thirty students who will ask "but why?" until the answer is real exposes every gap in your own mental model. That habit transfers directly to ML work: if I can't explain why a model behaves the way it does, I don't ship it.
Your stakeholders are students
Presenting model results to clinicians and research leads is the same skill as teaching: meet people where they are, strip jargon, and build from what they already know. The data scientist who can explain a confusion matrix to a physiotherapist is worth more than one who can only explain it to other data scientists.
Marking teaches you to debug
Grading hundreds of assignments trains you to spot the exact line where reasoning went wrong — fast. It's the same muscle as reviewing code or diagnosing a training run that won't converge.
If you get the chance to teach — a workshop, a lunch-and-learn, a tutorial — take it. Your engineering will thank you.