Prostate cancer is a heterogeneous disease, displaying a multitude of genetic alterations, histological patterns and clinical outcomes. This heterogen

点击翻译按钮获取中文摘要

The main aims of this project is to leverage recent advances in Natural Language Processing (NLP) to develop end-to-end clinical support systems which can utilise longitudinal free text documents within Electronic Health Records (EHRs). EHRs will often contain historic records, pertaining to all interactions between a patient and the healthcare service, including freetext documents, such as referral letters and discharge notes. A notable challenge is being able to adequately capturing longitudinal representations of clinical texts. Common state-of-the-art models such as the Bidirectional Encoder Representations fromTransformers (BERT) can only process sequences of 512 tokens (Devlin et al., 2018), but a years worth of clinical text for a single patient can consist of more than 10, 000 tokens. Another more general problem relates to the transparency, interpretability and algorithmic fairness of large language models. Therefore this project aims to develop methods and protocol to enhance these aspects.One proposed approach to representing sequential free-textbuilds upon the signature of a path, a non-parametric approach to extracting features from data in the form of tensors(Chevyrev and Kormilitzin, 2016). Loosely speaking, a signature is a collection of statistics about a stream of data that are time invariant, and has universal non-linearity, whereby it is sufficient to capture all possibly nonlinear functions of the original data: allowing aunique approach to representing complex sequential data. Combining signature techniques with strategies to address the limited ability of attention mechanisms in common transformer basedmodels, such as spare-attention mechanisms (Zaheer et al., 2020). This hybrid approach should allow efficient computation and representations of patients clinical text history, usable in a numberof relevant downstream tasks.Another approach will embrace a new paradigm shift in NLP research, named prompt-learning. Traditional approaches tomany downstream tasks involved taking a model such as BERT pre-trained on masked language modelling (MLM) and next sentence prediction (NSP) followed by a fine-tuning process on downstream tasks. Prompt-learning instead reconstructs the pretraining to embed the downstream task, encouraging the model to implicitly learn the desired task. The use of prompt-learning in a clinical domain has not been documented yet, thus provides a great opportunity.The proposed new methodologies will be developed and implemented in consultation with clinicians and will address real clinical use-cases. Specifically, the language models will be trained on a large collection of free-text notes from secondary care UKCRIS database to help triage patients to specialist teams. Other strands will explore the feasibility of identifying patients for clinical trials and identification of self-harm. The feasibility of translation of the developed methodology and models will be tested beyond the scope of mental health under the support provided by the EPSRC CDT in Health Data Science.This project falls within the EPSRC healthcare technologies research area.