Machine Learning

Transfer Learning

A method for measuring similarity in time series datasets for predicting transfer learning performance

 One of the most interesting research areas in deep learning is transfer learning. By pre-training a network on a good, reliable dataset, that network can then be retrained for the purpose of another task, where data is limited. If transfer learning is successful, then the pre-trained dataset will have greater performance than a network only trained on the data-limited dataset. In order for transfer learning to work, there needs to be some similarity between the dataset used for pre-training (source dataset) and the data-limited dataset (target dataset). When the datasets are dissimilar, negative transfer learning may occur, which lowers the performance of the target dataset by confusing the network with useless information. We are developing a method of measuring the similarity of two datasets, for use in predicting if transfer learning will provide a boost in performance between two datasets. The type of data we are looking at is time series signal data.

Our method is being tested by searching for a suitable time series dataset for use in a study in which the goal is to predict the stage of recovery a youth participant may be in after being diagnosed with a concussion. The prediction is being made from only the participants movement throughout the study, by use of an accelerometer sensor. This dataset is data-limited, so transfer learning is a good candidate to increase the performance of a model trained on this dataset. the similarity metric developed will be used to search for a suitable source dataset for transfer learning.

Principal investigators

Carol DeMatteo, Professor, Faculty of Health Sciences, McMaster University

Michael Noseworthy, Professor, School of Biomedical Engineering, McMaster University

Researchers

Ryan Clark, MASc Candidate

Dr. Thomas E. Doyle, PhD PEng, Associate Professor, Department of Electrical and Computer Engineering, McMaster University, School of Biomedical Engineering, McMaster University, Vector Institute

Cluster Analysis

A Clustering Approach to Identify Groupings among Chronic Pain Conditions

Pain is considered a major global healthcare problem and it is one of the most prevalent causes for patients seeking medical attention. Global estimates suggest that 1 in 10 adults are diagnosed with chronic pain. Though chronic pain and its associated diseases are not immediately life-threatening, they can have serious effects on a person's daily life including depression, anxiety, inability to work, disruption in social relationships, and suicidal thoughts. According to the Canadian Pain Task Force Report of 2019, many Canadians lack access to appropriate pain management services that can lead to poor treatment in the early stages, and it exacerbates with time. So, the need for generalization and optimization of pain treatments is crucial to offer the best possible treatments with limited resources.

Identifying the underlying cause is the most effective way of managing and treating chronic pain. While the number of potential treatment targets has increased and mechanism-based along with individualized pain therapy approaches are introduced, chronic pain is still very challenging to manage clinically. Usually, these conditions are divided into subtypes by clinical examinations based on their anatomical location and relevant symptoms overlooking etiology that might hamper the optimum treatment or management. Additionally, these empirical classification approaches are limited in scope and often disregard known pathophysiological mechanisms which consequently leads to sub-optimal treatment outcomes. These can also add cognitive bias in the diagnosis affecting the patient outcomes. Thus, there is a need for an approach to identify the pain phenotypes based on the exact mechanisms driving them while eliminating the effect of bias to improve chronic pain treatment and management.

This work intends to use data of patients with chronic pain and try to identify trends that have clinically meaningful groupings of patients using unsupervised learning or clustering. These approaches would help analyze the data regardless of the diagnosis established at the pain clinic. The objective is if we can identify these groupings, we may improve AI algorithms across other pain groups/cohorts. Here, a challenge would be to add reasoning or causal analysis in the algorithm and to make it explainable to people from other domains. Also, to ensure, inject, and build trust while maintaining ethics, reliability, transparency into the core of the hypothesized approach is another important aspect to be addressed. If pain can be clustered into different groups or sub-types other than the classical anatomical approach then that could lead to optimized management and resource allocation for hospitals, pain clinics, and rehabilitation centers. This can also be helpful for the development of personalized rehabilitation programs. In addition to that, it will open the opportunity to increase effectiveness and satisfaction among the patients which is significant in pain management.

Student Principal Investigator

Md Asif Khan, MASc Student, Department of Electrical and Computer Engineering, McMaster University.

Supervisor

Dr. Thomas E. Doyle, PhD PEng, Associate Professor, Department of Electrical and Computer Engineering, McMaster University, School of Biomedical Engineering, McMaster University, Vector Institute

Co-supervisor

Dr. Dinesh Kumbhare, Associate Professor & Clinician Scientist, Department of Medicine, Physical Medicine and Rehabilitation, University of Toronto, Canada.

Healthcare & Injury