Talks

Title:  Integrating Knowledge Graph Data with Large Language Models for Explainable Inference
Speakers: Carlos Efraín Quintero Narvaez
Abstract:

Recent advancements in Large Language Models (LLMs) such as OpenAI GPT, BERT, and LLaMA have demonstrated their potential for complex reasoning tasks with natural language. However, there is still room for improvement as it is costly to train these models to work with specialized data, and their inner workings are not yet fully understood. In this line of thought, Neurosymbolic Artificial Intelligence, which combines Symbolic Logic Reasoning and Deep Learning, aims to create explainable inference models using the virtues of the two fields. Knowledge Graphs (KGs) are an essential component in this subject, since they provide concise representations of large knowledge bases, understandable for both users and models. Two significant challenges in this area are query answering from KGs, and the integration of KG information into the output of language models. To address these issues, researchers have proposed various approaches, including the use of Deep Learning for complex queries on KGs and Augmented Language Models that integrate recognition of entities from a KG. In this thesis, we propose to modify and combine these approaches with recent LLM developments, creating an explainable way for LLMs to work with data from any KG. Our approach will use LLMs for the KG entity embedding steps utilized in existing techniques, while keeping the other parts of the methods intact. Furthermore, we will use this new querying method for executing KG queries in the internal inference process of LLMs. Using an architecture that integrates entity embeddings to the model’s inference. Our goal is to reduce the frequency of hallucinations and enhance the coherency of LLMs by allowing them to provide informed explanations, making them more broadly useful for general contexts.

We will be running a joint workshop with members of FHAIVE, Dario Greco's research group based in Tampere, Finland. We will explore some of the applications of AI in the biomedical domain.

Local Organiser: Paola Galdi

The Programme is as follows:

Title:  Exploring AI Based Approaches for Post-operative Microvascular Free Flap Monitoring
Speakers: Fiona Smith
Abstract:

Microvascular free flap surgery is a key technique utilised for the reconstruction of complex tissue deficits secondary to trauma and cancer. Whilst overall success rates are very good, this is only achieved because of close postoperative monitoring that allows the prompt identification of complications that require immediate re-operation for salvage. The current gold standard for free flap monitoring is regular clinical assessment of the patient at the beside but this is subjective, very labour intensive and disruptive for patients’ sleep. This PhD project aims to explore AI-based approaches for the postoperative monitoring of microvascular free flaps that are as effective and accurate as clinical assessment at identifying compromised free flaps but which are less subjective.  In this talk I will discuss the progress that has been made in the first year towards this aim. Firstly, I will outline the project protocols that have been designed for the acquisition of a free flap database with approvals for AI work. Secondly, I will describe the scoping review of current practices for ethical AI analysis of medical image datasets that was prompted by this. Finally, the results of early experiments to use a convolutional neural network to segment images of free flaps will be discussed and the plans for future work will be given.

Title:  The Usage and Improvement of Neurosymbolic AI for Biomedical Applications
Speakers: Lauren DeLong
Abstract:

Neurosymbolic AI describes a hybrid field of AI between symbolic methods, which tend to be robust and interpretable, and deep learning methods, which are more scalable and tend to perform more competitively. Since these two areas tend to see tradeoffs between the types of benefits they offer, methods in neurosymbolic AI try to leverage the perks of each while mitigating their respective weaknesses.  As neurosymbolic AI has recently increased in popularity, many various approaches have been introduced, and several of these approaches possess unique features that neither symbolic nor deep learning methods had alone. Consequently, neurosymbolic methods have potential to provide novel solutions to longstanding research challenges. The goal of my thesis is to use and improve neurosymbolic AI to demonstrate how these unique characteristics are especially useful for challenges which are particularly prominent in biomedical data science. Specifically, I will discuss challenges like meaningful multimodal data  integration for clinical datasets, discovering drug mechanisms of action via long-range dependencies, and few shot learning to predict and understand rare, serious side effects.

Title:  Predicting in-hospital mortality for ICU patients with liver disease using process mining and deep learning
Speakers: Zonglin Ji
Abstract:

Intensive Care Unit (ICU) patients with liver disease present a unique clinical challenge due to increased risks of complications and mortality. Traditional severity scoring methods, while helpful, do not comprehensively consider a patient's prior medical history nor the temporal events associated with their stay, and are thus limited when it comes to providing an up-to-date understanding of a patient’s condition as it evolves. This research aims to incorporate time series events from the care pathways of patients to improve on this situation and enable better survival predictions over time. In this talk, we describe our work on combining Process Mining and Deep Learning and applying it to mortality predictions for ICU patients with liver diseases.

Title:  Process-aware pattern recognition and anomaly detection under uncertainty
Speakers: Jiawei Zheng
Abstract:

In today's increasingly digital world, recognising patterns and detecting anomalies have become increasingly crucial for ensuring efficient operations, enhancing productivity and promoting human healthy. Processes often embody contextual information, such as domain-specific knowledge, optimal or expected behaviour, or established best practices. By considering the process information from context, we can interpret patterns more accurately and enhance the performance of anomaly detection. However, the structure of underlying processes in different domains varies greatly. Due to this, the patterns and anomalies we are interested in have different focuses and come from different perspectives, and thus pose different challenges in different domains.

In this talk, I will present the problems associated with incorporating process knowledge for pattern recognition and anomaly detection over uncertain data in heterogenous domains, the current progress that I have achieved, and finally the future plans to finish the project.

Title: Probabilistic Resources
Speakers: Filip Smola
Abstract:

When composing processes, non-deterministic resources allow us to express uncertainty such as decisions, sensing or possible failure. However, in practice we want to know not only the possible cases but also their relative probabilities. We may use these to compute the expected value of some performance metric or to uncover impossible execution paths.

In this talk I will present our work on adding these probabilities to a theory of process composition mechanised in Isabelle/HOL. I will give a high-level overview of how the probabilistic information is represented, the issues we encounter and how we address those.

Title:  Algebraic Formalisation with locales, types and relations
Speakers: Richard Schmoetten
Abstract:

Different approaches have been explored when formalising mathematical structures in Isabelle/HOL, roughly split between the set-based approach using locales and type-based classes. We present our formal theories in Lie groups and real algebra to introduce both ideas, and examine a relation-based way of reasoning about algebraic morphisms in a type class. This has the advantage of integrating directly with automated tooling in Isabelle (the transfer package), and may present an alternative to the current trend of recasting type-based theories as set-based ones (e.g. the Types-to-Sets package).

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.