Digital health monitoring 


Not all walking is the same. Changes in the way we walk, talk or move, can be indicative of our health or disease progression. Traditional medical examination provides only a snippet in ones health. Wearable devices can be used to measure how our walking, voice or quality of life changes continiously and this can empower medical experts with a newfound understanding of a range of conditions such as Parkinson's disease. However, deriving robust and meaningful digital biomarkers for Parkinson's is difficult open problem.

Published work:

  • Automatic quality control and enhancement for voice-based remote Parkinson’s disease detection, Speech Communication

  • Probabilistic modelling of gait for robust passive monitoring in daily life, IEEE Journal of Biomedical and Health Informatics

  • Real Life Gait Performance as Biomarker of Motor fluctuations, Journal of medical Internet research

  • Probabilistic modelling for unsupervised analysis of human behaviour in smart cities, Sensors

  • Quality control of voice recordings in remote Parkinson’s disease monitoring using the infinite hidden Markov model, ICASSP

  • Automated quality control for sensor-based symptom measurement performed outside the lab, Sensors

In peer review:

  • Few-shot time series segmentation using prototype-defined infinite hidden Markov models 

Media coverage on Parkinson's monitoring:

What to do when K-means fails? 


Despite the plethora of more flexible approaches to clustering, simple techniques such as variants of K-means remain  widely used due to their scalability and interpretable behaviour. Exploring links betweek K-means and graphical models, we have developed flexible alternatives of K-means which address many of its problems related to: (1) setting the number of clusters; (2) initialization (3)  cluster sphericity and density; (4) scalability. We also develop hierarchical extensions related to clustering across different contexts. 

Published work:

  • What to Do When K-Means Clustering Fails: A Simple yet Principled Alternative Algorithm, PLoS One

  • Simple approximate MAP inference for Dirichlet processes mixtures, Electronic Journal of Statistics

  • A deterministic inference framework for discrete nonparametric latent variable models: learning complex probabilistic models with simple algorithms, PhD Thesis

Finding shared factors and subspaces in high dimensional data


Doing inference in high-dimensional spaces is one of the core challenges in statistical machine learning. A workhorse in addressing this challenge are linear dimensionality techniques such as principal component analysis and factor analysis. Here we look at principled piecewise linear methods to dimensionality reduction which achieve both sparser decomposition and identify clustering structure in the original space.

Adaptive probabilistic principal component analysis, BNP@Neurips

Controlling for sparsity in sparse factor analysis models:adaptive latent feature sharing for piecewise linear dimensionality reduction

Can you monitoring the occupancy of a room using a single motion sensor? 


Occupancy estimation systems are becoming ubiquitous part of smart buildings. However, majority of them rely on privacy invading camera sensors or multiple motion sensors placed at all exits. We show that approximate occupancy estimation can be easily achieved using a novel affordable setup with a single motion sensor.

Published work:

Single-cell RNA sequence clustering 


In this study, we compare a large set of popular single-cell clustering algorithms and pre-processing strategies. The study demonstrates that the clustering performance of single-cell data often depends on just as much on appropriate preprocessing, as it does on the specific clustering algorithm. We also develop a novel Bayesian nonparametric method for single-cell clustering with the ensMAPDP package release coming soon.

  • ​Benchmark and parameter sensitivity analysis of scRNAseq clustering methods, Frontiers in Genetics​​