Talks

Title:  Standards for Reporting Artificial Intelligence Research using Burns Image Datasets
Speakers: Fiona Smith
Abstract:

In recent years there has been increasing interest in AI applications in Plastic Surgery. In particular, the additional objectivity that computer vision approaches could bring to otherwise subjective visual analysis has been identified as being of particular utility to a speciality which concerns itself with the restoration of form and function. The literature includes several systematic reviews of proof-of-concept studies and early applications of computer vision approaches in the Plastic Surgery domain but their conclusions have often been hindered by a lack of standardised reporting of methods and results.  This review firstly explores the current legislative environment for AI research that uses patient image datasets and then summarises key ethical considerations that are raised on review of AI burns care research. Finally, a basic framework for the reporting of burns image datasets that are used for AI research is suggested. It is hoped that this work will contribute to wider legislative discussions with more stakeholders.

Title: Drug Mechanism of Action Retrieval using Neurosymbolic Path Finding
Speakers: Lauren DeLong
Abstract:

Researchers I will be collaborating with have discovered novel plant-based drug compounds that possess some therapeutic effect. However, they do not yet understand how these compounds achieve such effects, otherwise known as the drug's mechanism of action (MOA). Uncovering drug MOAs is non-trivial: while some drug compounds act directly upon a cellular component which induces some therapeutic effect, many others act indirectly through a series of interactions. Therefore, the goal of my internship project is to use neurosymbolic AI on a biomedical knowledge graph to unveil the most likely paths by which these compounds execute their MOAs. By revealing the most likely MOAs, my method will assist the drug discovery process by facilitating a deeper understanding of: (1) how directly or indirectly a given drug compound achieves some therapeutic effect,  (2) what additional therapeutic or adverse effects can be expected from usage of said compound, and (3) how such effects may vary across cell and tissue types.

Title: Charting the Landscape of Neuro-symbolic Reasoners
Speakers: Xuelong An
Abstract:

In the first half of my presentation, I want to share an ongoing work on building a comprehensive benchmark to empirically assess the plethora of neuro-symbolic models. We note that in recent years, interest over this family of models is growing, evidenced by the constant influx of novel methods and benchmarks to test their robust generalization and reasoning capabilities. However, much of the successes reported by neuro-symbolic methods over assessed datasets are often disparate with respect to one another. There lacks a unified, comprehensive test to assess the panorama of neuro-symbolic models. To design such benchmark, we survey the current landscape of neuro-symbolic architectures and benchmarks. From this, we propose a general taxonomy for classifying current and future neuro-symbolic models and reasoning benchmarks, which helps us understand how they relate to each other. Henceforth, we propose SaSSY-CLEVR, a heterogeneous benchmark suite which can serve as a common testing ground for different neuro-symbolic reasoners to compare their strengths and limitations.

If time allows, in the second half of my presentation, I will share a series of experiments to assess NeSy models on CLEVR-Hans3, which test for the ability of object-centric reasoning adopted in SaSSY-CLEVR. In our study, we expand on work done by Stammer et. al (2021),  where we test the robustness of a traditional convolutional neural networks (CNN) and Neuro-Symbolic (NeSy) architectures comprising of a Slot Attention and a Set Transformer component. We evaluate different NeSy variants by comparing their classification accuracy after fine-tuning them to a modified version of the CLEVR-Hans3 dataset containing four different kinds of data complications. We find that models using the pretrained Slot Attention maintained good classification performance across data complications, indicating that the object-centric representations built by this perceptual component are crucial for model robustness. We also find that a Slot Attention with fully connected layers, instead of a Set Transformer, had the best overall performance, underscoring the importance of controlled comparisons.

Title: Knowledge-graph approaches to predict adverse events from electronic health records
Speakers: Paola Galdi
Abstract:

With older age, there is an increased chance of being diagnosed with more than one long-term condition. The medical treatment of patients with multiple conditions is challenging because the interactions of symptoms and medications are complex and hard to predict. In this talk, I will discuss an ongoing project using knowledge-graph methods to detect people who are likely to have unexpected health problems (like falls or bleeding), with the ultimate goal of supporting doctors in the choice of proper treatment and preventive care. I will briefly introduce the Clinical Practice Research Datalink (CPRD) dataset and the data model underlying the knowledge graph. I will then present a first attempt at repurposing a knowledge graph recommender system (KGAT) in the context of adverse events predictions. I will conclude with an overview of the challenges and open questions left to address.

Title: Constrained Training of Neural Networks via Theorem Proving
Speakers: Mark Chevallier and Matt Whyte
Abstract:

Robotic movement can be trained using neural networks but the process can be lengthy and has no guarantee that safety rules are learned. A neurosymbolic approach can bring the benefits of formal logical constraints specifying safety rules that can be injected into the training process. We introduce our work using logical constraints to assist the training of neural networks via a theorem proving process. In addition to learning via imitation, the neural network evaluates and learns from error caused by breaching these constraints. Our process formally proves the soundness of the logical loss function and guarantees correct implementation of that function using code generation. We discuss our existing work using linear temporal logic to train dynamic movement primitives and go on to discuss future extensions.

Title: ML-based premise selection for Lean
Speaker: Ramon Fernández Mir
Abstract:

In this talk, I will introduce a machine-learning-based tool for the Lean theorem prover that suggests relevant premises to a user interactively constructing a proof. The tool, entirely written in Lean 4, is designed to be highly user-friendly, customizable, and efficient. It is based on a version of random forest, trained on data extracted from mathlib -- Lean's mathematics library. I will discuss one of the main challenges, which was producing useful training features and labels. Finally, I will give a short demo and talk about some interesting related work.

Title: Understanding the Rehabilitation Pathways in Hospital for Patients with Acute COVID-19: A Process Mining Approach
Speaker: Konstantin Georgiev
Abstract:

The delivery of in-hospital rehabilitation during the pandemic presented a serious cause for concern. Patients surviving the infection stage frequently required prolonged and more complex treatment. Thus, rehabilitation wards have undergone rapid changes in the delivery of these services. For the first time, detailed contact data with associated timestamps and activities performed by Allied Healthcare Professionals (AHPs) has become routinely available in Electronic Health Records. This data could be key to understanding complex rehabilitation pathways and identifying areas for improvement in COVID-19, resulting in better clinical outcomes. In this study, I will explore the sequences of AHP-related activities in patients presenting with COVID-19 across three acute hospitals within NHS Lothian. I will try to assess the treatment efficiency of the delivered services between the first and second waves using process maps and metrics relative to recovery time.

Title: Stochastic block modelling and link prediction to improve mortality prediction for critical patients.
Speaker: Jorge Gaete Villegas
Abstract:

Mortality prediction for patients in Intensive Care Units (ICU) is an important but challenging task. Early prediction can improve medical outcomes, optimize medical interventions, and minimize the use of resources. Current efforts to create mortality prediction models rely on medical consensus, regression methods, and machine learning. Unfortunately, the nature and quality of  ICU data can affect the performance of such models. Some of the shortcomings reported in the literature include the overestimation of mortality for older patients and low predictive power for underrepresented patient groups. In this talk we present our current work exploring stochastic block modelling and link prediction to forecast mortality and overcome such shortcomings.

Title: What do data tell us about frailty?
Speaker: Lara Johnson
Abstract:

My research integrates data science and geriatric medicine to explore what different data sources and methods can tell us about frailty, a state of increased vulnerability to adverse health outcomes for individuals of the same chronological age. I am looking at how the number and combination of health issues people have – a proxy measure for frailty - relate to their functioning ability (such as their ability to make a cup of tea, get dressed or walk up a flight of stairs) and adverse health outcomes (death, falls, fractures, care needs).  This will inform the development of a data-driven definition of frailty (currently lacking), which is useful both for identifying patients in later life at highest risk as well as forecasting demand on health and social care services. My research aims to answer questions such as whether there are different types of frailty, whether frailty manifests differently in men vs. women (who are more frail but live longer) and whether distinguishing between types of frailty improves the performance of prediction models.

Title: Human activity recognition and Identifying the activity patterns on noninvasive sensor data with Deep Learning
Speaker: Simon U
Abstract:
The recent advancement and development of both embedded electronic devices and deep learning techniques have made real-time activity tracking and monitoring feasible with the help of wearable devices or cameras. The time-series data can be analyzed and used to track an individual's health condition and daily activities routine. But there's a growing trend toward using noninvasive and non-visual activity sensing to get information and figure out what a person is doing without bothering them. No one wants to be constantly watched and recorded by cameras.

The current approach for human activity recognition using ambient sensors, such as motion sensors and light sensors, is restrictive and has poorer performance compared to approaches using cameras and wearable sensors. In this talk, I will discuss the challenges encountered on the task as well as potential approaches I hope to investigate in order to overcome some of these issues during my MInf project. The aim is to build a model that takes into account prior knowledge of common activity patterns for the activity recognition task and uses the same model to forecast the individualized activity routine.