Quantitative analysis of social signals in neurological disorders

Living in an aging society means caring for an increasing number of people with neurological illnesses such as Parkinson's and Alzheimher's, but it also offers unprecendented opportunities to address these illnesses technologically. This project, with Alex 'Sandy' Pentland (Human Dynamics Group, MIT Media Lab) and funded in part by the Wellcome Trust, Intel Corporation, and the UK EPSRC, aims to radically lower the cost and inconvenience of clinical monitoring of the symptoms of neurological diseases by exploiting knowledge of social signalling, physiology, non-invasive measurements, distributed communications such as the internet and mobile devices, and advanced signal processing algorithms. These algorithms produce a simple measure of the symptom severity on a standard clinical scale used by doctors (the Unified Parkinson's Disease Rating Scale), and together with my student Athanasios Tsanas at Oxford, we have demonstrated that it is possible to predict, from non-invasive speech recordings, Parkinson's symptoms on the UPDRS scale with a few precent error (figure shows remote symptom tracking over the duration of a clinical trial).

Nonlinear signal processing of piecewise constant signals

The mathematical foundations of linear, time-invariant systems that underpin classical signal processing are elegant and well understood. However, not all signals are optimally processed this way. For instance, if a signal contains noise but has jumps, then any linear filter that can remove the noise, also must smooth away jumps to a certain extent. Nonlinear signal processing offers practical solutions in this surprisingly common circumstance. This project introduced a new framework for the problem, and developed many new nonlinear signal processing algorithms for noise removal from piecewise constant signals (figure shows the effectiveness of the new mean-shift total variation algorithm).

Physical model-based analysis tools for molecular machines

Nanotechnology has nothing on nature. For example, most bacteria have a "tail" that is used for locomotion to seek out nutrients and avoid hazards, and this tail is rotated by a fully functioning nano-scale stepper motor, powered by ions produced in the mitochondria. Novel experimental assays can track the rotation of these motors and other molecular machines in vivo. But these experimental techniques produce a lot of noisy data, and analysis of this data is computationally intensive, and often fails to produce "crisp" results across all the experimental data. This project with Nick Jones (Oxford Complex Systems and Oxford Centre for Integrative Systems Biology) and Richard Berry has developed novel data analysis tools that can find robust signatures of the step-like dynamics of these machines, by building in physical models of the molecular dynamics. Using these tools, we have pinpointed robust signatures of the symmetry and mechanochemical coupling of different components of the E-coli flagellar motor (figure shows robust extraction of rotational motor symmetries from signals of motor angle against time).

Forecasting regional climate changes in extreme rainfall

This is the anthropocene, and the climate is changing rapidly. Climate models of the atmosphere agree that global surface temperatures will increase, but they disagree about predicted changes in patterns of extreme rainfall at the regional level, for example, in the UK. This project, originally funded by the NERC, aims to predict regional changes in extreme rainfall patterns using advanced statistical time series analysis and machine learning methods (figure shows areas in the UK where extreme rainfall has increased by 20% - blue - and decreased by 20% - orange - since 1961). We have been able to show that our statistical models outperform state-of-the-art ensemble numerical weather models when predicting regional rainfall.