Personalized online ensemble machine learning with applications for dynamic data streams.

In this work we introduce the personalized online super learner (POSL), an online personalizable ensemble machine learning algorithm for streaming data. POSL optimizes predictions with respect to baseline covariates, so personalization can vary from completely individualized, that is, optimization with respect to subject ID, to many individuals, that is, optimization with respect to common baseline covariates. As an online algorithm, POSL learns in real time. As a super learner, POSL is grounded in statistical optimality theory and can leverage a diversity of candidate algorithms, including online algorithms with different training and update times, fixed/offline algorithms that are not updated during POSL's fitting procedure, pooled algorithms that learn from many individuals' time series, and individualized algorithms that learn from within a single time series. POSL's ensembling of the candidates can depend on the amount of data collected, the stationarity of the time series, and the mutual characteristics of a group of time series. Depending on the underlying data-generating process and the information available in the data, POSL is able to adapt to learning across samples, through time, or both. For a range of simulations that reflect realistic forecasting scenarios and in a medical application, we examine the performance of POSL relative to other current ensembling and online learning methods. We show that POSL is able to provide reliable predictions for both short and long time series, and it's able to adjust to changing data-generating environments. We further cultivate POSL's practicality by extending it to settings where time series dynamically enter and exit.

Statistics, philosophy, and health: the SMAC 2021 webconference.

SMAC 2021 was a webconference organized in June 2021. The aim of this conference was to bring together data scientists, (bio)statisticians, philosophers, and any person interested in the questions of causality and Bayesian statistics, ranging from technical to philosophical aspects. This webconference consisted of keynote speakers and contributed speakers, and closed with a round-table organized in an unusual fashion. Indeed, organisers asked world renowned scientists to prepare two videos: a short video presenting a question of interest to them and a longer one presenting their point of view on the question. The first video served as a "teaser" for the conference and the second were presented during the conference as an introduction to the round-table. These videos and this round-table generated original scientific insights and discussion worthy of being shared with the community which we do by means of this paper.

Precision machine learning to understand micro-RNA regulation in neurodegenerative diseases.

Micro-RNAs (miRNAs) are short (∼21 nt) non-coding RNAs that regulate gene expression through the degradation or translational repression of mRNAs. Accumulating evidence points to a role of miRNA regulation in the pathogenesis of a wide range of neurodegenerative (ND) diseases such as, for example, Alzheimer's disease, Parkinson's disease, amyotrophic lateral sclerosis and Huntington disease (HD). Several systems level studies aimed to explore the role of miRNA regulation in NDs, but these studies remain challenging. Part of the problem may be related to the lack of sufficiently rich or homogeneous data, such as time series or cell-type-specific data obtained in model systems or human biosamples, to account for context dependency. Part of the problem may also be related to the methodological challenges associated with the accurate system-level modeling of miRNA and mRNA data. Here, we critically review the main families of machine learning methods used to analyze expression data, highlighting the added value of using shape-analysis concepts as a solution for precisely modeling highly dimensional miRNA and mRNA data such as the ones obtained in the study of the HD process, and elaborating on the potential of these concepts and methods for modeling complex omics data.

Post-Contextual-Bandit Inference.

Contextual bandit algorithms are increasingly replacing non-adaptive A/B tests in e-commerce, healthcare, and policymaking because they can both improve outcomes for study participants and increase the chance of identifying good or even best policies. To support credible inference on novel interventions at the end of the study, nonetheless, we still want to construct valid confidence intervals on average treatment effects, subgroup effects, or value of new policies. The adaptive nature of the data collected by contextual bandit algorithms, however, makes this difficult: standard estimators are no longer asymptotically normally distributed and classic confidence intervals fail to provide correct coverage. While this has been addressed in non-contextual settings by using stabilized estimators, the contextual setting poses unique challenges that we tackle for the first time in this paper. We propose the Contextual Adaptive Doubly Robust (CADR) estimator, the first estimator for policy value that is asymptotically normal under contextual adaptive data collection. The main technical challenge in constructing CADR is designing adaptive and consistent conditional standard deviation estimators for stabilization. Extensive numerical experiments using 57 OpenML datasets demonstrate that confidence intervals based on CADR uniquely provide correct coverage.

Risk Minimization from Adaptively Collected Data: Guarantees for Supervised and Policy Learning.

Empirical risk minimization (ERM) is the workhorse of machine learning, whether for classification and regression or for off-policy policy learning, but its model-agnostic guarantees can fail when we use adaptively collected data, such as the result of running a contextual bandit algorithm. We study a generic importance sampling weighted ERM algorithm for using adaptively collected data to minimize the average of a loss function over a hypothesis class and provide first-of-their-kind generalization guarantees and fast convergence rates. Our results are based on a new maximal inequality that carefully leverages the importance sampling structure to obtain rates with the good dependence on the exploration rate in the data. For regression, we provide fast rates that leverage the strong convexity of squared-error loss. For policy learning, we provide regret guarantees that close an open gap in the existing literature whenever exploration decays to zero, as is the case for bandit-collected data. An empirical investigation validates our theory.

Prediction of an Acute Hypotensive Episode During an ICU Hospitalization With a Super Learner Machine-Learning Algorithm.

BACKGROUND: Acute hypotensive episodes (AHE), defined as a drop in the mean arterial pressure (MAP) <65 mm Hg lasting at least 5 consecutive minutes, are among the most critical events in the intensive care unit (ICU). They are known to be associated with adverse outcome in critically ill patients. AHE prediction is of prime interest because it could allow for treatment adjustment to predict or shorten AHE. METHODS: The Super Learner (SL) algorithm is an ensemble machine-learning algorithm that we specifically trained to predict an AHE 10 minutes in advance. Potential predictors included age, sex, type of care unit, severity scores, and time-evolving characteristics such as mechanical ventilation, vasopressors, or sedation medication as well as features extracted from physiological signals: heart rate, pulse oximetry, and arterial blood pressure. The algorithm was trained on the Medical Information Mart for Intensive Care dataset (MIMIC II) database. Internal validation was based on the area under the receiver operating characteristic curve (AUROC) and the Brier score (BS). External validation was performed using an external dataset from Lariboisière hospital, Paris, France. RESULTS: Among 1151 patients included, 826 (72%) patients had at least 1 AHE during their ICU stay. Using 1 single random period per patient, the SL algorithm with Haar wavelets transform preprocessing was associated with an AUROC of 0.929 (95% confidence interval [CI], 0.899-0.958) and a BS of 0.08. Using all available periods for each patient, SL with Haar wavelets transform preprocessing was associated with an AUROC of 0.890 (95% CI, 0.886-0.895) and a BS of 0.11. In the external validation cohort, the AUROC reached 0.884 (95% CI, 0.775-0.993) with 1 random period per patient and 0.889 (0.768-1) with all available periods and BSs <0.1. CONCLUSIONS: The SL algorithm exhibits good performance for the prediction of an AHE 10 minutes ahead of time. It allows an efficient, robust, and rapid evaluation of the risk of hypotension that opens the way to routine use.

Performance Guarantees for Policy Learning.

This article gives performance guarantees for the regret decay in optimal policy estimation. We give a margin-free result showing that the regret decay for estimating a within-class optimal policy is second-order for empirical risk minimizers over Donsker classes when the data are generated from a fixed data distribution that does not change with sample size, with regret decaying at a faster rate than the standard error of an efficient estimator of the value of an optimal policy. We also present a result giving guarantees on the regret decay of policy estimators for the case that the policy falls within a restricted class and the data are generated from local perturbations of a fixed distribution, where this guarantee is uniform in the direction of the local perturbation. Finally, we give a result from the classification literature that shows that faster regret decay is possible via plug-in estimation provided a margin condition holds. Three examples are considered. In these examples, the regret is expressed in terms of either the mean value or the median value, and the number of possible actions is either two or finitely many.

Cet article présente des garanties de performance concernant la vitesse à laquelle le regret s'amenuise dans le cadre de l'estimation d'une politique d'action optimale. Si la politique optimale est définie comme optimale relativement à un ensemble de politiques formant une classe de Donsker, et si elle est estimée par minimisation sur cet ensemble d'une estimation du regret vu comme une fonction sur celui-ci, alors un premier résultat révèle que la vitesse est de second ordre dès lors que les observations sont générées sous une loi qui ne change pas à mesure que leur nombre augmente. Plus spécifiquement, le regret de l'estimateur de la politique optimale s'amenuise plus rapidement que l'écart type d'un estimateur efficace de la valeur d'une politique optimale. Ce résultat ne nécessite pas le recours à une hypothèse de marge. Un second résultat porte sur la vitesse à laquelle le regret de l'estimateur de la politique optimale s'amenuise lorsque les observations sont générées sous des lois définies comme des perturbations locales d'une loi de référence fixe, la garantie de performance étant alors uniforme relativement aux directions de perturbation. Finalement, un troisième résultat montre qu'il est possible d'atteindre des vitesses plus rapides en mettant en œuvre une procédure d'estimation par substitution à la condition qu'une hypothèse de marge soit satisfaite. Ce résultat s'inspire de la littérature consacrée à la classification. Trois exemples illustrent nos trouvailles. Dans ceux-ci, le regret s'exprime en termes de valeur moyenne ou de valeur médiane, et les actions envisageables sont au nombre de deux ou bien en nombre fini.

Scalable collaborative targeted learning for high-dimensional data.

Robust inference of a low-dimensional parameter in a large semi-parametric model relies on external estimators of infinite-dimensional features of the distribution of the data. Typically, only one of the latter is optimized for the sake of constructing a well-behaved estimator of the low-dimensional parameter of interest. Optimizing more than one of them for the sake of achieving a better bias-variance trade-off in the estimation of the parameter of interest is the core idea driving the general template of the collaborative targeted minimum loss-based estimation procedure. The original instantiation of the collaborative targeted minimum loss-based estimation template can be presented as a greedy forward stepwise collaborative targeted minimum loss-based estimation algorithm. It does not scale well when the number p of covariates increases drastically. This motivates the introduction of a novel instantiation of the collaborative targeted minimum loss-based estimation template where the covariates are pre-ordered. Its time complexity is O(p) as opposed to the original O(p2) , a remarkable gain. We propose two pre-ordering strategies and suggest a rule of thumb to develop other meaningful strategies. Because it is usually unclear a priori which pre-ordering strategy to choose, we also introduce another instantiation called SL-C-TMLE algorithm that enables the data-driven choice of the better pre-ordering strategy given the problem at hand. Its time complexity is O(p) as well. The computational burden and relative performance of these algorithms were compared in simulation studies involving fully synthetic data or partially synthetic data based on a real world large electronic health database; and in analyses of three real, large electronic health databases. In all analyses involving electronic health databases, the greedy collaborative targeted minimum loss-based estimation algorithm is unacceptably slow. Simulation studies seem to indicate that our scalable collaborative targeted minimum loss-based estimation and SL-C-TMLE algorithms work well. All C-TMLEs are publicly available in a Julia software package.

Big data and targeted machine learning in action to assist medical decision in the ICU.

Historically, personalised medicine has been synonymous with pharmacogenomics and oncology. We argue for a new framework for personalised medicine analytics that capitalises on more detailed patient-level data and leverages recent advances in causal inference and machine learning tailored towards decision support applicable to critically ill patients. We discuss how advances in data technology and statistics are providing new opportunities for asking more targeted questions regarding patient treatment, and how this can be applied in the intensive care unit to better predict patient-centred outcomes, help in the discovery of new treatment regimens associated with improved outcomes, and ultimately how these rules can be learned in real-time for the patient.

TARGETED SEQUENTIAL DESIGN FOR TARGETED LEARNING INFERENCE OF THE OPTIMAL TREATMENT RULE AND ITS MEAN REWARD.

This article studies the targeted sequential inference of an optimal treatment rule (TR) and its mean reward in the non-exceptional case, i.e., assuming that there is no stratum of the baseline covariates where treatment is neither beneficial nor harmful, and under a companion margin assumption. Our pivotal estimator, whose definition hinges on the targeted minimum loss estimation (TMLE) principle, actually infers the mean reward under the current estimate of the optimal TR. This data-adaptive statistical parameter is worthy of interest on its own. Our main result is a central limit theorem which enables the construction of confidence intervals on both mean rewards under the current estimate of the optimal TR and under the optimal TR itself. The asymptotic variance of the estimator takes the form of the variance of an efficient influence curve at a limiting distribution, allowing to discuss the efficiency of inference. As a by product, we also derive confidence intervals on two cumulated pseudo-regrets, a key notion in the study of bandits problems. A simulation study illustrates the procedure. One of the corner-stones of the theoretical study is a new maximal inequality for martingales with respect to the uniform entropy integral.

tmle.npvi: targeted, integrative search of associations between DNA copy number and gene expression, accounting for DNA methylation.

UNLABELLED: We describe the implementation of the method introduced by Chambaz et al. in 2012. We also demonstrate its genome-wide application to the integrative search of new regions with strong association between DNA copy number and gene expression accounting for DNA methylation in breast cancers. AVAILABILITY AND IMPLEMENTATION: An open-source R package tmle.npvi is available from CRAN (http://cran.r-project.org/). CONTACT: pierre.neuvial@genopole.cnrs.fr.

Analysis of the effect of occupational exposure to asbestos based on threshold regression modeling of case-control data.

We analyze the effect of occupational exposure to asbestos on the occurrence of lung cancer based on a recent French case-control (CC) study. We build a large collection of threshold regression models, data-adaptively select a better model by CC-weighted likelihood-based cross-validation and then fit this better model by CC-weighted maximum likelihood. The CC-weighting allows us to draw valid inferences from CC data without relying on a logistic regression. This is possible because the joint distribution of the indicator of being a case and matching variable is available beforehand owing to two studies independent from our data set. The implications of the fitted model in terms of years of life free of lung cancer lost due to the exposure to asbestos are discussed.

Estimation of a non-parametric variable importance measure of a continuous exposure.

We define a new measure of variable importance of an exposure on a continuous outcome, accounting for potential confounders. The exposure features a reference level x(0) with positive mass and a continuum of other levels. For the purpose of estimating it, we fully develop the semi-parametric estimation methodology called targeted minimum loss estimation methodology (TMLE) [23, 22]. We cover the whole spectrum of its theoretical study (convergence of the iterative procedure which is at the core of the TMLE methodology; consistency and asymptotic normality of the estimator), practical implementation, simulation study and application to a genomic example that originally motivated this article. In the latter, the exposure X and response Y are, respectively, the DNA copy number and expression level of a given gene in a cancer cell. Here, the reference level is x(0) = 2, that is the expected DNA copy number in a normal cell. The confounder is a measure of the methylation of the gene. The fact that there is no clear biological indication that X and Y can be interpreted as an exposure and a response, respectively, is not problematic.

Control of neuronal persistent activity by voltage-dependent dendritic properties.

Neural integrators and working memory rely on persistent activity, a widespread neural phenomenon potentially involving persistent sodium conductances. Using a unique combination of voltage-clamp, dynamic-clamp, and frequency-domain techniques, we have investigated the role of voltage-dependent conductances on the dendritic electrotonic structure of neurons of the prepositus hypoglossi nucleus (PHN), which is known to be involved in oculomotor integration. The PHN contains two main neuronal populations: type B neurons with a double afterhyperpolarization and type D neurons, which not only are oscillatory but also have a greater electrotonic length than that of type B neurons. The persistent sodium conductance is present in all PHN neurons, although its effect on the dynamic electrotonic structure is shown to significantly differ in the two major cell types present in the nucleus. The electrotonic differences are such that the persistent sodium conductance can be almost perfectly manipulated in a type B neuron using an on-line dynamic clamp to add or subtract virtual sodium ion channels. The dynamic-clamp results are confirmed by data-fitted models, which suggest that the persistent sodium conductance has two different roles depending on its somatic versus dendritic location: perisomatic conductances could play a major role in maintaining action potential discharge and dendritic conductances would be more involved in other computational properties, such as those involving remote synaptic processing or bistable events.

Plica semilunaris temporal ectopia: evidence of primary nasal pterygia traction.

PURPOSE: To study whether plica semilunaris (PS) temporal ectopia is associated with primary nasal pterygia and whether such hypothetical ectopia is linked to with-the-rule astigmatism so that a discussion of the clinical and surgical implications of such possible findings might take place. METHODS: Cross-section observational study of 208 consecutive eyes of 152 cases and 169 eyes of 109 control subjects enrolled for 6 months. Four PS position grades were designed. Grade 1 features a grossly vertical PS position. Grades 2 and 3 involve temporal displacement and an abnormal PS tilt. Grade 4 involves PS obliteration with possible associated temporal caruncle dragging. Pterygium corneal encroachment, pterygium fleshiness, and manifest refraction were recorded. RESULTS: PS position was significantly displaced temporally in pterygium-affected eyes (P = 2 x 10(-4)). Corneal encroachment (P = 2.1 x 10(-5)), pterygium fleshiness (P = 2.7 x 10(-5)), and age (P = 1.3 x 10(-2)) were positively correlated with PS position grades > 1. The presence of with-the-rule astigmatism was significantly linked to PS position grades > 1 (P = 0.01). CONCLUSIONS: Primary nasal pterygia result in PS temporal ectopia, and PS position grades (more than grade 1) are linked to the presence of with-the-rule astigmatism. These findings are consistent with the exertion of tractional forces at both pterygium ends. In eye abduction, medial canthus structures (PS, caruncule) may be dragged temporally. Thus, pterygium-related corneal traction may not always significantly increase in eye abduction. PS temporal ectopia may influence the decision to remove PS during pterygium surgical excision.

Sifting abstracts from Medline and evaluating their relevance to molecular biology.

The most important knowledge in the area of biology currently consists of raw text documents. Bibliographic databases of biomedical articles can be searched, but an efficient procedure should evaluate the relevance of documents to biology. In genetics, this challenge is even trickier, because of the lack of consistency in genes' naming tradition. We aim to define a good approach for collecting relevant abstracts for biology and for studied species and genes. Our approach relies on defining best queries, detecting and filtering best sources.