For the end of our Summer School, we offered participants from all backgrounds to present their own work, studies, or company to the community. Here is the list of the talks that were given:
Jia Le Xian
Challenges of real world impact: our lessons learned in creating an automatic treatment system for Covid-19
Challenges and our lessons learned in the process of bringing a machine learning powered stay home treatment system for mild cases of Covid -19 into clinical settings.
Fine grained sleep apnea tagging with deep learning on unselected patients using routinely collected vital signs
AIOSA is a model that aims to tag with 1-second granularity apnea events to estimate obstructive sleep apnea syndrome severity on unselected semi-intensive ward patients. We will briefly discuss our approach and the advantages of working with such a granularity. Then, we will remark the challenges and the implications related to our specific, real-world setting (OSASUD dataset): unselected patients, relying only routinely recorded vital signs, working in electrically hostile environments (stroke units).
Interpretable Multi-Modal Sleep Monitoring System using Ear-EEG and EOG
Despite improved performance, Black-box behavior of deep-learning algorithms have limited their use in the clinical settings. Our recent work focuses on developing an interpretable sleep stage classification algorithm using Transformers.
link to paper : https://arxiv.org/abs/2208.06991
link to code: https://github.com/Jathurshan0330/Cross-Modal-Transformer
RSI Foundation – a nonprofit organization to help people deal with chronic pain using AI
1 – Patient empowerment through information to counteract maladaptive behavior and stigmatization, specifically with respect to work-related pain and the central component of chronic pain
2 – Develop free and accessible Natural Language Processing-based tools to enable hands-free computer use
3 – Raise interest in pain research, particularly in the AI community, to foster collaboration and interdisciplinary communication
Dr Ghazanfar Khan
AI for good : A global health perspective
How can AI and ML be used to help solve the global public health challenges of our time? Join this 5 minute quick fire talk on challenges and opportunities for adoption in health systems.
How Roche is using Data and Advanced Analytics
Roche has access to a wide variety of biomedical data of varying types from a diverse range of sources, and are investigating a range of ML & AI approaches to utilise these data to improve our clinical trials and develop tools which will improve patient care.
Applications of AI in Provider and Population Health Management
1 – Automated Provider Services Mapping — NPHIES codes (over 40k codes) were introduced in a bid to unify and standardize services across the health sector. Every service provider is therefore expected to comply. To ensure compliance with the regulation, we consequently developed an NLP (Siamese Neural Network) based human-assisted mapping system to speed up the process.
2 – Chronic Diseases Prediction – an AI based model to identify members to be enrolled into proactive intervention programs in order to stall or totally halt the progression of certain chronic diseases.
Prediction of Missense Variant Effects on Quantitative Traits From Population Cohort Data
Regression models that can accurately predict the impact of missense variants from exome sequencing data on phenotypes as measured in over 200’000 patients in the UK Biobank. These models elucidate variants with moderate monogenic effects within established trait-gene relationships, which can provide information useful for candidate screening and biological investigation.
Clinical presence: how does the interaction between patients and healthcare systems impact our machine learning pipelines?
Observational data are the results of the interaction between patients and the healthcare system. This interaction results in an informative data-generative process, often ignored in machine learning pipelines. Our work explores the impact of this process on pipelines’ transportability and fairness.
Visual Streak Localization in Spectral Domain Optical Coherence Tomography Images of Minipigs
OCT data is often noisy and has lots of missing points. Our approach uses the Bayesian hierarchical model to denoise and to fill missing values. Then we apply scale-dependent decomposition and credibility analysis to localize the visual streak, an area that has higher photoreceptor density in minipig’s retina.
Fuzzy k-plane clustering method with local spatial information for segmentation of human brain MRI image.” Neural Computing and Applications
In this research work, we incorporated local spatial information in the FkPC clustering method to handle the noise present in the image. This spatial regularization term included in the proposed FkPCS method refines the membership value of a noisy pixel with the help of immediate neighbour pixels information. To show the effectiveness of the proposed FkPCS method, extensive experiments are performed on one synthetic image and two publicly available human brain MRI datasets.
Dr. Alejandro Diaz
CHIMERA (Collaborative Health Innovation through Mathematics, Engineering and AI)
CHIMERA is one of four national Hubs for Mathematical Sciences in Healthcare funded by EPSRC, based at UCL. We are examining data from intensive care unit patients through collaborative research between mathematical modelling, data science and AI.
Deep Approximation of Retinal Traits (DART) for robust and efficient retinal image analysis
Retinal images show a detailed picture of the retinal vasculature and are thus not only of interest for eye diseases but also systemic vascular conditions like cardiovascular disease. Existing retinal image analysis pipelines require high image quality (~25-40% of the data in datasets like UK Biobank needs to be discarded) and are computationally expensive (processing all images in UK Biobank can take weeks or months). We propose Deep Approximation of Retinal Traits (DART) where we learn to approximate an existing pipeline with a deep neural network while injecting robustness to common image quality issues through data augmentation. DART can process over 1,000 images per second (all of UK Biobank takes less than an hour on a single machine) and recover retinal image traits from severely degraded images that correlate well with the original measure
Raul Sena Rojas
SMOTE with Active Learning
It’s a research work at PSL University about using active learning with SMOTE to deal with the class imbalance in medical datasets.
Latent-Variable Random Forest improves feature importance interpretability by considering features’ neighbourhood
We developed a Latent-Variable Random Forest model which highlights the local neighbourhood of features, hence improving the feature importance interpretability and is unaffected by single ‘important’ noisy features (false positives).
Efficient feature selection using sparse neural networks
Major complications arise from the recent increase in the amount of high-dimensional data, including high computational costs and memory requirements. Feature selection, which identifies the most relevant and informative attributes of a dataset, has been introduced as a solution to this problem. Most of the existing feature selection methods are computationally inefficient; inefficient algorithms lead to high energy consumption, which is not desirable for devices with limited computational and energy resources. In this talk, I will present how we will use sparse neural networks to perform efficient feature selection.
Palgun N P
The project is to Classify Alzheimer’s disease based on the MRI scans into its level of severity. It also aims to perform Brain Tumour Segmentation. This is achieved using CNNs
Dr Angèle Bénard-Sankaran
Empowering patients with standardised outcomes data and insights to drive healthcare improvements
H2O: Building European data infrastructure for health outcomes.
BioModels ML – building an open repository of FAIR ML models in life science and medicine
We are extending BioModels repository (https://www.ebi.ac.uk/biomodels) to build an open collection of findable, accessible, interoperable, and reusable (FAIR) and reproducible machine learning models in life science and healthcare.