Language Based Decision Support System For Treatment Planning In Patients With Sub-arachnoid Haemorrhage.

Sub-arachnoid haemorrhage (or bleeding over the surface of the brain,) is a debilitating disease that strikes suddenly and without warning. Until recently the prospects for sufferers of this illness were very bleak, with death in almost half and severe disability in most survivors. The United Kingdom has pioneered advanced treatments for this disease including new keyhole-surgery techniques which offer much greater survival and greatly reduced disability. However these new techniques are very complex and are being developed at such a fast rate, that the clinicians looking after the patients are often unsure as to which exact form of treatment to offer. In addition it is often very difficult to give patients and their relatives exact information on their chances of survival as well as the type and risks of the disabilities they may suffer. This project will design a computer system that analyses data about individual patients and their scans, using this information to guide the doctor through the decision making process. It will update itself taking into account new research as it becomes available and keep track of how each individual patient has done. In this way it will become a world reference point for doctors treating sub-arachnoid haemorrhage, giving up to the minute accurate treatment advice tailored for individual patients and offering real-time prognosis information. This system will keep the United Kingdom at the forefront in the treatment of this often catastrophic illness.

Aims: Develop and evaluate the core components of a language-based decision support system (DSS) for therapeutic planning in patients with aneurysmal sub-arachnoid haemorrhage (SAH). Objectives: Produce a descriptive language model forming the basis of a Bayesian DSS to guide treatment decisions, specifically: 1) compile a list of descriptive features which have shown prognostic implications for outcome; 2) assess feature inter and intra-observer reproducibility; 3) examine computer-based guides for improving baseline reproducibility; 4) identify the prognostic value of individual features for a range of clinical outcomes; 5) develop a web based reporting system using a Bayesian DSS algorithm and simple language model, enabling prognostic classification of new cases in real-time. Design: Prospective studies such as the International Sub-arachnoid Aneurysm Trial (ISAT) have confirmed the benefits of endovascular versus surgical treatment however there is remarkably little data to guide therapeutic decisions. Some imaging features are predominant in treatment planning whereas others have a lack of objective evidence. Imaging data will be used to generate probabilities against a series of clinical outcomes allowing predictions to be derived using Bayes theory. The proposed method based on Bayesian DSS is an important study on how to combine expert knowledge with machine learning (learning the posterior probabilities based on Bayesian rules), which is an important and fundamental issue in knowledge discovery and decision support. Methodology: Imaging feature descriptive terms will be identified from a review of world literature. The ability of neuroradiologists to identify these features will be assessed to determine which features suffer from inter and intra observer error. The features which can be reproduced will be compared to known patient outcomes to determine their initial prognostic values. A DSS framework and language model will then be implemented allowing predictive outcome probabilities to be determined using Bayes theory. Validation will then be performed using 100 patients from the original ISAT trial not used previously in the project. Scientific and medical opportunities: The project will generate extensive evidence concerning the independent value of predictive features in SAH, and provide a substantial feasibility study of a DSS to support ongoing multicentre studies of SAH therapy. It represents the core of a system which could be improved by expansion of outcome indicators, determinant features and integration of improved techniques for feature identification such as formalised ontologies, natural language models or automated image analysis.