What is a KTP?
A Knowledge Transfer Partnership (KTP) is a UK government-funded scheme that connects a university with a business to work on a strategic innovation project. An academic team provides the research expertise, the company provides the real-world problem and data, and an embedded associate — me — sits between the two, managing the project day-to-day and transferring the knowledge into the business.
The Project
A partnership between the University of Hertfordshire and Transpharmation. We're building a facial-video analysis platform: starting with pain-detection proof-of-concepts, then extending towards measuring disease state and treatment response — Parkinson's, stroke and dementia. The aim is to read clinically meaningful signals from ordinary video, with no contact sensors.
My Role
- Own the project end-to-end, bridging the university and the company.
- Drive the machine-learning research: pipelines, models, evaluation.
- Compare sensing modalities (video vs physiological signals) to find what actually carries the signal.
- Run experiments at scale on the university HPC cluster.
Progress Log
Project meeting: presented noisy-subject removal and video feature reduction. Ranking the 912-D facial features by importance and keeping only the top ~30 actually improved results — around 86% with a leave-one-subject-out random forest, and EDA alone reaching ~79% versus the ~72% video-only baseline from the original study. The findings are now strong enough to write up: a paper on the results is on the way, targeting a journal special issue.
Project meeting on the multimodal comparison. Across SVM (linear and RBF) and random forest on BioVid: EDA (skin conductance) is the strongest single signal for telling graded pain levels apart, while the video features give the best and most stable accuracy for simply detecting whether pain is present. Feature-level fusion of the signals and the challenge of individual variability were the main discussion points.
Ran the first proper multimodal comparison on the BioVid dataset: the existing 912-D facial-landmark video features vs physiological signals (EDA, ECG, EMG), with SVM and Random Forest under leave-one-subject-out cross-validation on the university HPC cluster. Goal: quantify how much each sensing modality actually contributes to pain classification.
Built a side-research webapp to understand remote photoplethysmography (rPPG) — recovering heart rate from an ordinary face video. Useful background for the contactless physiological-signal angle of the project.
Pain-detection proof-of-concept pipelines taking shape: data ingestion, feature extraction and a first classification baseline end-to-end.
Started the role: Research Scientist on a Knowledge Transfer Partnership between the University of Hertfordshire and Transpharmation, a pharmaceutical research company. I manage the project end-to-end, bridging the university and the company and driving the machine-learning research.