How AI can learn from the law: putting humans in the loop only on appeal.

While the literature on putting a "human in the loop" in artificial intelligence (AI) and machine learning (ML) has grown significantly, limited attention has been paid to how human expertise ought to be combined with AI/ML judgments. This design question arises because of the ubiquity and quantity of algorithmic decisions being made today in the face of widespread public reluctance to forgo human expert judgment. To resolve this conflict, we propose that human expert judges be included via appeals processes for review of algorithmic decisions. Thus, the human intervenes only in a limited number of cases and only after an initial AI/ML judgment has been made. Based on an analogy with appellate processes in judiciary decision-making, we argue that this is, in many respects, a more efficient way to divide the labor between a human and a machine. Human reviewers can add more nuanced clinical, moral, or legal reasoning, and they can consider case-specific information that is not easily quantified and, as such, not available to the AI/ML at an initial stage. In doing so, the human can serve as a crucial error correction check on the AI/ML, while retaining much of the efficiency of AI/ML's use in the decision-making process. In this paper, we develop these widely applicable arguments while focusing primarily on examples from the use of AI/ML in medicine, including organ allocation, fertility care, and hospital readmission.

Pandemic lockdown, isolation, and exit policies based on machine learning predictions.

The widespread lockdowns imposed in many countries at the beginning of the COVID-19 pandemic elevated the importance of research on pandemic management when medical solutions such as vaccines are unavailable. We present a framework that combines a standard epidemiological SEIR (susceptible-exposed-infected-removed) model with an equally standard machine learning classification model for clinical severity risk, defined as an individual's risk of needing intensive care unit (ICU) treatment if infected. Using COVID-19-related data and estimates for France as of spring 2020, we then simulate isolation and exit policies. Our simulations show that policies considering clinical risk predictions could relax isolation restrictions for millions of the lowest risk population months earlier while consistently abiding by ICU capacity restrictions. Exit policies without risk predictions, meanwhile, would considerably exceed ICU capacity or require the isolation of a substantial portion of population for over a year in order to not overwhelm the medical system. Sensitivity analyses further decompose the impact of various elements of our models on the observed effects. Our work indicates that predictive modeling based on machine learning and artificial intelligence could bring significant value to managing pandemics. Such a strategy, however, requires governments to develop policies and invest in infrastructure to operationalize personalized isolation and exit policies based on risk predictions at scale. This includes health data policies to train predictive models and apply them to all residents, as well as policies for targeted resource allocation to maintain strict isolation for high-risk individuals.

Computational ethics.

Technological advances are enabling roles for machines that present novel ethical challenges. The study of 'AI ethics' has emerged to confront these challenges, and connects perspectives from philosophy, computer science, law, and economics. Less represented in these interdisciplinary efforts is the perspective of cognitive science. We propose a framework - computational ethics - that specifies how the ethical challenges of AI can be partially addressed by incorporating the study of human moral decision-making. The driver of this framework is a computational version of reflective equilibrium (RE), an approach that seeks coherence between considered judgments and governing principles. The framework has two goals: (i) to inform the engineering of ethical AI systems, and (ii) to characterize human moral judgment and decision-making in computational terms. Working jointly towards these two goals will create the opportunity to integrate diverse research questions, bring together multiple academic communities, uncover new interdisciplinary research topics, and shed light on centuries-old philosophical questions.

The European artificial intelligence strategy: implications and challenges for digital health.

In February, 2020, the European Commission published a white paper on artificial intelligence (AI) as well as an accompanying communication and report. The paper sets out policy options to facilitate a secure and trustworthy development of AI and considers health to be one of its most important areas of application. We illustrate that the European Commission's approach, as applied to medical AI, presents some challenges that can be detrimental if not addressed. In particular, we discuss the issues of European values and European data, the update problem of AI systems, and the challenges of new trade-offs such as privacy, cybersecurity, accuracy, and intellectual property rights. We also outline what we view as the most important next steps in the Commission's iterative process. Although the European Commission has done good work in setting out a European approach for AI, we conclude that this approach will be more difficult to implement in health care. It will require careful balancing of core values, detailed consideration of nuances of health and AI technologies, and a keen eye on the political winds and global competition.

The need for a system view to regulate artificial intelligence/machine learning-based software as medical device.

Artificial intelligence (AI) and Machine learning (ML) systems in medicine are poised to significantly improve health care, for example, by offering earlier diagnoses of diseases or recommending optimally individualized treatment plans. However, the emergence of AI/ML in medicine also creates challenges, which regulators must pay attention to. Which medical AI/ML-based products should be reviewed by regulators? What evidence should be required to permit marketing for AI/ML-based software as a medical device (SaMD)? How can we ensure the safety and effectiveness of AI/ML-based SaMD that may change over time as they are applied to new data? The U.S. Food and Drug Administration (FDA), for example, has recently proposed a discussion paper to address some of these issues. But it misses an important point: we argue that regulators like the FDA need to widen their scope from evaluating medical AI/ML-based products to assessing systems. This shift in perspective-from a product view to a system view-is central to maximizing the safety and efficacy of AI/ML in health care, but it also poses significant challenges for agencies like the FDA who are used to regulating products, not systems. We offer several suggestions for regulators to make this challenging but important transition.

Reproducible evaluation of classification methods in Alzheimer's disease: Framework and application to MRI and PET data.

A large number of papers have introduced novel machine learning and feature extraction methods for automatic classification of Alzheimer's disease (AD). However, while the vast majority of these works use the public dataset ADNI for evaluation, they are difficult to reproduce because different key components of the validation are often not readily available. These components include selected participants and input data, image preprocessing and cross-validation procedures. The performance of the different approaches is also difficult to compare objectively. In particular, it is often difficult to assess which part of the method (e.g. preprocessing, feature extraction or classification algorithms) provides a real improvement, if any. In the present paper, we propose a framework for reproducible and objective classification experiments in AD using three publicly available datasets (ADNI, AIBL and OASIS). The framework comprises: i) automatic conversion of the three datasets into a standard format (BIDS); ii) a modular set of preprocessing pipelines, feature extraction and classification methods, together with an evaluation framework, that provide a baseline for benchmarking the different components. We demonstrate the use of the framework for a large-scale evaluation on 1960 participants using T1 MRI and FDG PET data. In this evaluation, we assess the influence of different modalities, preprocessing, feature types (regional or voxel-based features), classifiers, training set sizes and datasets. Performances were in line with the state-of-the-art. FDG PET outperformed T1 MRI for all classification tasks. No difference in performance was found for the use of different atlases, image smoothing, partial volume correction of FDG PET images, or feature type. Linear SVM and L2-logistic regression resulted in similar performance and both outperformed random forests. The classification performance increased along with the number of subjects used for training. Classifiers trained on ADNI generalized well to AIBL and OASIS. All the code of the framework and the experiments is publicly available: general-purpose tools have been integrated into the Clinica software (www.clinica.run) and the paper-specific code is available at: https://gitlab.icm-institute.org/aramislab/AD-ML.