Synthetic data generation is one approach for sharing individual-level data. However, to meet legislative requirements, it is necessary to demonstrate that the individuals' privacy is adequately protected. There is no consolidated standard for measuring privacy in synthetic data. Through an expert panel and consensus process, we developed a framework for evaluating privacy in synthetic data. Our findings indicate that current similarity metrics fail to measure identity disclosure, and their use is discouraged. For differentially private synthetic data, a privacy budget other than close to zero was not considered interpretable. There was consensus on the importance of membership and attribute disclosure, both of which involve inferring personal information about an individual without necessarily revealing their identity. The resultant framework provides precise recommendations for metrics that address these types of disclosures effectively. Our findings further present specific opportunities for future research that can help with widespread adoption of synthetic data.

This thesis aims to study some of the mathematical challenges that arise in the analysis of statistical sequential decision-making algorithms for postoperative patients follow-up. Stochastic bandits (multiarmed, contextual) model the learning of a sequence of actions (policy) by an agent in an uncertain environment in order to maximise observed rewards. To learn optimal policies, bandit algorithms have to balance the exploitation of current knowledge and the exploration of uncertain actions. Such algorithms have largely been studied and deployed in industrial applications with large datasets, low-risk decisions and clear modelling assumptions, such as clickthrough rate maximisation in online advertising. By contrast, digital health recommendations call for a whole new paradigm of small samples, risk-averse agents and complex, nonparametric modelling. To this end, we developed new safe, anytime-valid concentration bounds, (Bregman, empirical Chernoff), introduced a new framework for risk-aware contextual bandits (with elicitable risk measures) and analysed a novel class of nonparametric bandit algorithms under weak assumptions (Dirichlet sampling). In addition to the theoretical guarantees, these results are supported by in-depth empirical evidence. Finally, as a first step towards personalised postoperative follow-up recommendations, we developed with medical doctors and surgeons an interpretable machine learning model to predict the long-term weight trajectories of patients after bariatric surgery.

Investigators from Cedars-Sinai have created an artificial intelligence-enabled tool that may make it easier to predict if a person will have a heart attack. The tool, described in The Lancet Digital Health, accurately predicted which patients would experience a heart attack in five years based on the amount and composition of plaque in arteries that supply blood to the heart. Plaque buildup can cause arteries to narrow, which makes it difficult for blood to get to the heart, increasing the likelihood of a heart attack. A medical test called a coronary computed tomography angiography (CTA) takes 3D images of the heart and arteries and can give doctors an estimate of how much a patient's arteries have narrowed. Until now, however, there has not been a simple, automated and rapid way to measure the plaque visible in the CTA images.

A smart stethoscope combined with an AI algorithm could enable earlier diagnosis of heart failure and improved patient outcomes, all at reduced cost. Researchers at the National Heart and Lung Institute and Imperial's Centre for Cardiac Engineering have completed the first-ever NHS study to evaluate an artificial intelligence (AI) technology for point-of-care detection of heart failure. Heart failure is a condition in which the heart cannot pump blood effectively. It carries a higher risk of death than most cancers and is increasingly common, affecting 2% of the UK population and consuming 4% of the NHS budget. Reliably diagnosing heart failure is a major challenge for GPs as the symptoms, such as breathlessness are associated with many other conditions.

Prostate biopsy with cancer probability (blue is low, red is high). This case was originally diagnosed as benign but changed to cancer upon further review. The AI accurately detected cancer in this tricky case. A study published today (July 27, 2020) in The Lancet Digital Health by UPMC and University of Pittsburgh researchers demonstrates the highest accuracy to date in recognizing and characterizing prostate cancer using an artificial intelligence (AI) program. "Humans are good at recognizing anomalies, but they have their own biases or past experience," said senior author Rajiv Dhir, M.D., M.B.A., chief pathologist and vice chair of pathology at UPMC Shadyside and professor of biomedical informatics at Pitt. "Machines are detached from the whole story. To train the AI to recognize prostate cancer, Dhir and his colleagues provided images from more than a million parts of stained tissue slides taken from patient biopsies. Each image was labeled by expert pathologists to teach the AI how to discriminate between healthy and abnormal tissue. The algorithm was then tested on a separate set of 1,600 slides taken from 100 consecutive patients seen at UPMC for suspected prostate cancer. During testing, the AI demonstrated 98% sensitivity and 97% specificity at detecting prostate cancer -- significantly higher than previously reported for algorithms working from tissue slides. Also, this is the first algorithm to extend beyond cancer detection, reporting high performance for tumor grading, sizing, and invasion of the surrounding nerves. These all are clinically important features required as part of the pathology report. AI also flagged six slides that were not noted by the expert pathologists. But Dhir explained that this doesn't necessarily mean that the machine is superior to humans. For example, in the course of evaluating these cases, the pathologist could have simply seen enough evidence of malignancy elsewhere in that patient's samples to recommend treatment. For less experienced pathologists, though, the algorithm could act as a failsafe to catch cases that might otherwise be missed. "Algorithms like this are especially useful in lesions that are atypical," Dhir said. "A nonspecialized person may not be able to make the correct assessment.

Prostate biopsy with cancer probability (blue is low, red is high). This case was originally diagnosed as benign but changed to cancer upon further review. The AI accurately detected cancer in this tricky case. A study published today (July 27, 2020) in The Lancet Digital Health by UPMC and University of Pittsburgh researchers demonstrates the highest accuracy to date in recognizing and characterizing prostate cancer using an artificial intelligence (AI) program. "Humans are good at recognizing anomalies, but they have their own biases or past experience," said senior author Rajiv Dhir, M.D., M.B.A., chief pathologist and vice chair of pathology at UPMC Shadyside and professor of biomedical informatics at Pitt. "Machines are detached from the whole story. To train the AI to recognize prostate cancer, Dhir and his colleagues provided images from more than a million parts of stained tissue slides taken from patient biopsies. Each image was labeled by expert pathologists to teach the AI how to discriminate between healthy and abnormal tissue. The algorithm was then tested on a separate set of 1,600 slides taken from 100 consecutive patients seen at UPMC for suspected prostate cancer. During testing, the AI demonstrated 98% sensitivity and 97% specificity at detecting prostate cancer -- significantly higher than previously reported for algorithms working from tissue slides. Also, this is the first algorithm to extend beyond cancer detection, reporting high performance for tumor grading, sizing, and invasion of the surrounding nerves. These all are clinically important features required as part of the pathology report. AI also flagged six slides that were not noted by the expert pathologists. But Dhir explained that this doesn't necessarily mean that the machine is superior to humans. For example, in the course of evaluating these cases, the pathologist could have simply seen enough evidence of malignancy elsewhere in that patient's samples to recommend treatment. For less experienced pathologists, though, the algorithm could act as a failsafe to catch cases that might otherwise be missed. "Algorithms like this are especially useful in lesions that are atypical," Dhir said. "A nonspecialized person may not be able to make the correct assessment.

Researchers have developed an algorithm that can detect and identify different types of brain injuries. The team, from the University of Cambridge, Imperial College London and CONICET, have clinically validated and tested their method on large sets of CT scans and found that it was successfully able to detect, segment, quantify and differentiate different types of brain lesions. Their results, reported in The Lancet Digital Health, could be useful in large-scale research studies, for developing more personalised treatments for head injuries and, with further validation, could be useful in certain clinical scenarios, such as those where radiological expertise is at a premium. Head injury is a huge public health burden around the world and affects up to 60 million people each year. It is the leading cause of mortality in young adults.

Recently, Lancet Digital Health ran a meta-analysis concluding -- if cautiously -- that "deep learning" (more familiarly known as artificial intelligence) can be considered "equivalent to healthcare professionals" in image-based diagnoses. In an editorial commentary on the analysis, Tessa Cook says, in effect, "not so fast!"