Computer-vision based diagnosis of neurological diseases
The way we move our face call say a lot about our overall health. Facial movements, such as smile of patients with neurological conditions, such as Parkinson's disease, Alzheimer's disease, stroke, and amyotrophic lateral sclerosis, are different from those of age-matched healthy controls. Thus, analysis of facial movements can be used to diagnose, assess, and monitor treatment effectiveness in multiple clinical populations.
However, it is difficult to track how the face moves. Typically, specialized and expensive equipment, such as motion capture systems, is required to measure the movements of the face. Thus, this type of analysis is only performed in laboratory and research settings.
The main objective of this project is to apply and improve recent facial analysis development based on convolutional neural networks to create a clinically valid tool for analysis of facial movements. Our lab will develop and validate an open-source, easy-to-use software application that patients can use to evaluate their disease status objectively. Clinicians can also use this application to monitor the treatment effect from the patient’s home using a computer or mobile device.
Fairness in AI
Convolutional neural networks are typically trained with large datasets of labelled or unlabelled data. Collecting such data is expensive and time-consuming, and the datasets are often unbalanced, meaning that certain groups of categories are poorly represented in the data. Unbalanced dataset can have important negative consequences in the resulting models. For instance, facial recognition algorithms based on convolutional neural networks tend to be very accurate for white, young male subjects. However, the models’ accuracy when applied to a people of color is often much lower. There is a significant interest for developing balanced datasets to train fair algorithms, and companies such as Google and IBM are investing significant resources to create these new datasets. However, the problem is more challenging when dealing with clinical data, because collecting large datasets of clinical is often impossible due to the limited number of patients and privacy concerns.
In this project, we will explore how new machine learning techniques such as transfer learning and generative-adversarial networks can help to mitigate the problem of algorithmic bias against clinical populations without requiring large databases of representative data.
Assessmen of neuromuscular properties
Our joints define how we move and interact with the environment. These interactions often involve perturbations, such as stumbling, and joints must react to prevent damage, maintain balance, or preserve the movement trajectory. Voluntary reactions are not enough to prevent muscle damage or alteration joint trajectory produced by the perturbations, as they occur hundreds of milliseconds after the disturbance. Intrinsic and reflex stiffness are two additional, involuntary mechanisms that determine the instantaneous response of a joint to external perturbations. These mechanisms produce torques around the joint automatically in response to external perturbations. Thus, a comprehensive understanding of how the brain controls posture and movement must include a complete characterization of the joints’ intrinsic and reflex stiffness.
Intrinsic and reflex stiffness act and change together and their contributions to the overall joint stiffness cannot be measured directly. Thus, advanced biomechanical models, robotic interfaces, and systems identification techniques must be used to describe the intrinsic and reflex stiffness.
The main objectives of this project are to:
- Develop and validate a robotic interface to perturb human joints.
- Evaluate novel systems identification techniques to minimize data requirements and improve robustness against noise.
Identification of nonlinear, time-varying systems
Mathematical models are essential tools to describe and understand the behavior of physiological systems. Models provide a lot of information about the system under analysis and help to predict the effect of disease in the system function.
Most physiological systems cannot be modeled successfully as linear systems. And their response is influenced by external or internal factors. For example, the force produced by a muscle does not change linearly with the neural input; there are a minimum and maximum force that the muscle can produce. Furthermore, factors such as muscle length, velocity, fatigue, and previous activation greatly influence the force that the muscle can produce at a given time. Thus, models of physiological systems must account for these propeties of the systems.
This project aims to develop and validate novel analytical and experimental approaches to identify mathematical models used to describe physiological systems. In particular, we are interested in models that describe the human neuromuscular system.