An important part of law and regulation is demanding explanations for actual and potential failures. We ask questions like: What happened (or might happen) to cause this failure? And why did (or might) it happen? These are disguised normative questions -they really ask what ought to have happened, and how the humans involved ought to have behaved. If we ask the same questions about AI systems we run into two difficulties. The first is what might be described as the 'black box'problem, which lawyers have begun to investigate. Some modern AI systems are highly complex, so that even their makers might be unable to understand their workings fully, and thus answer the what and why questions. Technologists are beginning to work on this problem, aiming to use technology to explain the workings of autonomous systems more effectively, and also to produce autonomous systems which are easier to explain.But the second difficulty is so far underexplored, and is a more important one for law and regulation. This is that the kinds of explanation required by law and regulation are not, at least at first sight, the kinds of explanation which AI systems can currently provide.To answer the normative questions, law and regulation seeks a narrative explanation, a story. Humans usually explain their  Chris Reed, Professor of Electronic Commerce Law at the Centre for Commercial Law Studies,

Background Standard of care for locally advanced oesophageal adenocarcinoma is neoadjuvant chemotherapy or chemoradiotherapy followed by surgery. Only a minority of patients (<25%) derive significant survival benefit from neoadjuvant treatment and there are no reliable means of establishing prior to treatment in whom this benefit will occur. Moreover, accurate prediction of survival prior to treatment is also not possible. The availability of machine learning techniques provides the potential to use complex data sources to answer these problems. In this study, we assessed the utility of high-resolution digital microscopy of pre-treatment biopsies in predicting both response to neoadjuvant therapy and overall survival. Methods A total of 157 cases were included in the study. Pre-treatment clinical information, including neoadjuvant treatment, was obtained, along with diagnostic biopsies. Diagnostic biopsies were converted into high-resolution whole slide-images and features extracted using the pre-trained convolutional neural network Xception. Single representative images were converted into patches from which predictive models were trained. Elastic net regression classifiers were derived and validated with bootstrapping and 1000 resampled datasets. The response to treatment was considered according to Mandard tumour regression grade (TRG). Model performance was quantified using the C-index (for TRG) and time-dependent AUC (tAUC, fo Overall survival) along with calibration plots. Results Median survival was 78.9months (95%CI 35.9 months – not reached). Survival at 5-years was 52.1%. Neoadjuvant treatment was received by 123 patients (78.3%), with a significant response seen in 45 cases (36.6%). A response was more likely in those patients who received chemoradiotherapy than chemotherapy (53.3% vs 23.1% p < 0.001) and in older patients (median age 69.4 vs 66.0 years, p = 0.038), with other characteristics similar. A predictive model for response to neoadjuvant treatment derived from image features and clinical data achieved good discrimination (C-index 0.767, 95%CI 0.701-0.833) and calibration. Accuracy of prediction of overall survival was more modest (tAUC 0.640, 95%CI 0.518-0.762). Conclusions Using a small dataset, utility of a feature extraction pipeline in prediction of patient level outcomes has been demonstrated. This was more marked in prediction of response to neoadjuvant treatment than overall survival, which may reflect the importance of pre-treatment clinical data in determining the former outcome. Further study to refine the methodology and confirmation in larger datasets are required before expansion to clinical settings.

Introduction Locally advanced oesophageal adenocarcinoma is typically treated with neoadjuvant chemotherapy (NACT) or chemoradiotherapy (NACRT) followed by surgery. Significant benefit to neoadjuvant treatment however is confined to a minority of patients (<25%) and there are no reliable means of establishing prior to treatment in whom this benefit will occur. In this study, we assessed the utility of features extracted from high-resolution digital microscopy of pre-treatment biopsies in predicting response to neoadjuvant therapy in a machine-learning based modelling framework. Method A total of 102 cases were included in the study. Pre-treatment clinical information, including TNM staging, was obtained, along with diagnostic biopsies. Diagnostic biopsies were converted into high-resolution whole slide-images and features extracted using a pre-trained convolutional neural network (Xception). Elastic net regression models were then trained and validated with bootstrapping with 1000 resampled datasets. The response was considered according to Mandard tumour regression grade (TRG). Result There were 45 (44.1%) responders (TRG1-2) and 57 (57%) non-responders (TRG3-5) in the dataset. 34 patients (33.3%) received NACT and 68 (66.7%) received NACRT. A model trained with RNA-seq data achieved fair performance only in predicting response (AUC 0.598 95% CI 0.593–0.603), which was far exceeded by use of segmented diagnostic biopsy images (AUC 0.872 95% CI 0.869–0.875), which also produced well calibrated predictions of risk. Conclusion Despite using a small dataset, impressive performance in classifying response to neoadjuvant treatment can be achieved, particularly using automated image classification. Further study to refine the methodology is required before expansion to clinical settings. Take-home Message Response to neoadjuvant treatment for oesophageal cancer can be predicted from diagnostic biopsies