Legal Judgement Prediction (LJP) is the task of automatically predicting the outcome of a court case given only the case document. During the last five years researchers have successfully attempted this task for the supreme courts of three jurisdictions: the European Union, France, and China. Motivation includes the many real world applications including: a prediction system that can be used at the judgement drafting stage, and the identification of the most important words and phrases within a judgement. The aim of our research was to build, for the first time, an LJP model for UK court cases. This required the creation of a labelled data set of UK court judgements and the subsequent application of machine learning models. We evaluated different feature representations and different algorithms. Our best performing model achieved: 69.05% accuracy and 69.02 F1 score. We demonstrate that LJP is a promising area of further research for UK courts by achieving high model performance and the ability to easily extract useful features.

Objective To evaluate the performance of ASSIGN against the Framingham equations for predicting 10 year risk of cardiovascular disease in a UK cohort of patients from general practice and to make the evaluation comparable to an independent evaluation of QRISK on the same cohort. Design Prospective open cohort study. Setting 288 practices from England and Wales contributing to The Health Improvement Network (THIN) database. Participants Patients registered with 288 UK practices for some period between January 1995 and March 2006. The number of records available was 1 787 169. Main outcome measures First diagnosis of myocardial infarction, coronary heart disease, stroke and transient ischaemic attacks recorded. Methods We implemented the Anderson Framingham Coronary Heart Disease and Stroke models, ASSIGN, and a more recent Framingham Cox proportional-hazards model and analysed their calibration and discrimination. Results Calibration showed that all models tested over-estimated risk particularly for men. ASSIGN showed better discrimination with higher AUROC (0.756/0.792 for men/women), D statistic (1.35/1.58 for men/women), and R² (30.47%/37.39% for men/women). The performance of ASSIGN was comparable to that of QRISK on the same cohort. Models agreed on 93–97% of categorical (high/lower) risk assessments and when they disagreed, ASSIGN was often closer to the estimated Kaplan-Meier incidence. ASSIGN also provided a steeper gradient of deprivation and discriminated between those with and without recorded family history of CVD. The estimated incidence was twice/three times as high for women/men with a recorded family history of CVD. Conclusions For systematic CVD risk assessment all models could usefully be applied, but ASSIGN improved on the gradient of deprivation and accounted for recorded family history whereas the Framingham equations did not. However, all models display relatively low specificity and sensitivity. An additional conclusion is that the recording of family history of CVD in primary care databases needs to improve given its importance in risk assessment.

Missing data is a significant issue in many real-world datasets, yet there are no robust methods for dealing with it appropriately. In this paper, we propose a robust approach to dealing with missing data in classification problems: Multiple Imputation Ensembles (MIE). Our method integrates two approaches: multiple imputation and ensemble methods and compares two types of ensembles: bagging and stacking. We also propose a robust experimental set-up using 20 benchmark datasets from the UCI machine learning repository. For each dataset, we introduce increasing amounts of data Missing Completely at Random. Firstly, we use a number of single/multiple imputation methods to recover the missing values and then ensemble a number of different classifiers built on the imputed data. We assess the quality of the imputation by using dissimilarity measures. We also evaluate the MIE performance by comparing classification accuracy on the complete and imputed data. Furthermore, we use the accuracy of simple imputation as a benchmark for comparison. We find that our proposed approach combining multiple imputation with ensemble techniques outperform others, particularly as missing data increases.