A High Resolution Spatiotemporal Fine Particulate Matter Exposure Assessment Model for the Contiguous United States

Currently available nationwide prediction models for fine particulate matter (PM2.5) lack prediction confidence intervals and usually do not describe cross validated model performance at different spatiotemporal resolutions and extents. We used 41 different spatiotemporal predictors, including data on land use, meteorology, aerosol optical density, emissions, wildfires, population, traffic, and spatiotemporal indicators to train a machine learning model to predict daily averages of PM2.5 concentrations at 0.75 sq km resolution across the contiguous United States from 2000 through 2020. We utilized a generalized random forest model that allowed us to generate asymptotically-valid prediction confidence intervals and took advantage of its usefulness as an ensemble learner to quickly and cheaply characterize leave-one-location-out CV model performance for different temporal resolutions and geographic regions. Using a variable importance metric, we selected 8 predictors that were able to accurately predict daily PM2.5, with an overall leave-one-location-out cross validated median absolute error of 1.20 ug/m3, an R2 of 0.84, and confidence interval coverage fraction of 95%. When considering aggregated temporal windows, the model achieved leave-one-location-out cross validated median absolute errors of 0.99, 0.76, 0.63, and 0.60 ug/m3 for weekly, monthly, annual, and all-time exposure assessments, respectively. We further describe the model’s cross validated performance at different geographic regions in the United States, finding that it performs worse in the Western half of the country where there are less monitors. The code and data used to create this model are publicly available and we have developed software packages to be used for exposure assessment. This accurate exposure assessment model will be useful for epidemiologists seeking to study the health effects of PM across the continental United States, while possibly considering exposure estimation accuracy and uncertainty specific to the the spatiotemporal resolution and extent of their study design and population.

Temporal refinement of 3D CNN semantic segmentations on 4D time-series of undersampled tomograms using hidden Markov models

Recently, several convolutional neural networks have been proposed not only for 2D images, but also for 3D and 4D volume segmentation. Nevertheless, due to the large data size of the latter, acquiring a sufficient amount of training annotations is much more strenuous than in 2D images. For 4D time-series tomograms, this is usually handled by segmenting the constituent tomograms independently through time with 3D convolutional neural networks. Inter-volume information is therefore not utilized, potentially leading to temporal incoherence. In this paper, we attempt to resolve this by proposing two hidden Markov model variants that refine 4D segmentation labels made by 3D convolutional neural networks working on each time point. Our models utilize not only inter-volume information, but also the prediction confidence generated by the 3D segmentation convolutional neural networks themselves. To the best of our knowledge, this is the first attempt to refine 4D segmentations made by 3D convolutional neural networks using hidden Markov models. During experiments we test our models, qualitatively, quantitatively and behaviourally, using prespecified segmentations. We demonstrate in the domain of time series tomograms which are typically undersampled to allow more frequent capture; a particularly challenging problem. Finally, our dataset and code is publicly available.

Early prediction of hemodynamic interventions in the intensive care unit using machine learning

Background Timely recognition of hemodynamic instability in critically ill patients enables increased vigilance and early treatment opportunities. We develop the Hemodynamic Stability Index (HSI), which highlights situational awareness of possible hemodynamic instability occurring at the bedside and to prompt assessment for potential hemodynamic interventions. Methods We used an ensemble of decision trees to obtain a real-time risk score that predicts the initiation of hemodynamic interventions an hour into the future. We developed the model using the eICU Research Institute (eRI) database, based on adult ICU admissions from 2012 to 2016. A total of 208,375 ICU stays met the inclusion criteria, with 32,896 patients (prevalence = 18%) experiencing at least one instability event where they received one of the interventions during their stay. Predictors included vital signs, laboratory measurements, and ventilation settings. Results HSI showed significantly better performance compared to single parameters like systolic blood pressure and shock index (heart rate/systolic blood pressure) and showed good generalization across patient subgroups. HSI AUC was 0.82 and predicted 52% of all hemodynamic interventions with a lead time of 1-h with a specificity of 92%. In addition to predicting future hemodynamic interventions, our model provides confidence intervals and a ranked list of clinical features that contribute to each prediction. Importantly, HSI can use a sparse set of physiologic variables and abstains from making a prediction when the confidence is below an acceptable threshold. Conclusions The HSI algorithm provides a single score that summarizes hemodynamic status in real time using multiple physiologic parameters in patient monitors and electronic medical records (EMR). Importantly, HSI is designed for real-world deployment, demonstrating generalizability, strong performance under different data availability conditions, and providing model explanation in the form of feature importance and prediction confidence.

IDC: Quantitative Evaluation Benchmark of Interpretation Methods for Deep Text Classification Models

Recent advances in deep neural networks have achieved outstanding success in natural language processing. Due to the success and the black-box nature of the deep models, interpretation methods that provide insight into the decision-making process of the models have received an influx of research attention. However, there is no quantitative evaluation comparing interpretation methods for text classification other than observing classification accuracy or prediction confidence when important word grams are removed. This is due to the lack of interpretation ground truth. Manual labeling of a large interpretation ground truth is time-consuming. We propose IDC, a new benchmark for quantitative evaluation of I nterpretation methods for D eep text C lassification models. IDC consists of three methods that take existing text classification ground truth and generate three corresponding pseudo-interpretation ground truth datasets. We propose to use interpretation recall, interpretation precision, and Cohen’s kappa inter-agreement as performance metrics. We used the pseudo ground truth datasets and the metrics to evaluate six interpretation methods.

Dimensional control over metal halide perovskite crystallization guided by active learning

Metal halide perovskite (MHP) derivatives, a promising class of optoelectronic materials, have been synthesized with a range of dimensionalities that govern their optoelectronic properties and determine their applications. We demonstrate a data-driven approach combining active learning and high-throughput experimentation to discover, control, and understand the formation of phases with different dimensionalities in the morpholinium (morph) lead iodide system. Using a robot-assisted workflow, we synthesized and characterized two novel MHP derivatives that have distinct optical properties: a one-dimensional (1D) morphPbI3 phase ([C4H10NO][PbI3]) and a 2D (morph)2PbI4 phase ([C4H10NO]2[PbI4]). To efficiently acquire the data needed to construct a machine learning (ML) model of the reaction conditions where the 1D and 2D phases are formed, data acquisition was guided by a diverse-mini-batch-sampling active learning algorithm, using prediction confidence as a stopping criterion. Querying the ML model uncovered the reaction parameters that have the most significant effects on dimensionality control. Based on these insights, we propose a reaction scheme that rationalizes the formation of different dimensional MHP derivatives in the morph-Pb-I system. The data-driven approach presented here, including the use of additives to manipulate dimensionality, will be valuable for controlling the crystallization of a range of materials over large reaction-composition spaces.

A High Resolution Spatiotemporal Fine Particulate Matter Exposure Assessment Model for the Contiguous United States

Currently available nationwide prediction models for fine particulate matter (PM2.5) lack prediction confidence intervals and usually do not describe cross validated (CV) model performance at different spatiotemporal resolutions and extents. We used 41 different spatiotemporal predictors, including data on land use, meteorology, aerosol optical density, emissions, wildfires, population, traffic, and spatiotemporal indicators to train a machine learning model to predict daily averages of PM2.5 concentrations at 0.75 sq km resolution across the contiguous United States from 2000 through 2020. We utilized a generalized random forest model that allowed us to generate asymptotically-valid prediction confidence intervals and took advantage of its usefulness as an ensemble learner to quickly and cheaply characterize leave-one-location-out (LOLO) CV model performance for different temporal resolutions and geographic regions. Using a variable importance metric, we selected 8 predictors that were able to accurately predict daily PM2.5, with an overall LOLO CV median absolute error (MAE) of 1.20 μgm3, an R2 of 0.84, and confidence interval coverage fraction of 95%. When considering aggregated temporal windows, the model achieved LOLO CV MAEs of 0.99, 0.76, 0.63, and 0.60 μgm3 for weekly, monthly, annual, and all-time exposure assessments, respectively. We further describe the model’s CV performance at different geographic regions in the United States, finding that it performs worse in the Western half of the country where there are less monitors. The code and data used to create this model are publicly available and we have developed software packages to be used for exposure assessment. This accurate exposure assessment model will be useful for epidemiologists seeking to study the health effects of PM2.5 across the continential United States, while possibly considering exposure estimation accuracy and uncertainty specific to the the spatiotemporal resolution and extent of their study design and population.

Informing Pharmacokinetic Models With Physiological Data: Oral Population Modeling of L-Serine in Humans

To determine how to set optimal oral L-serine (serine) dose levels for a clinical trial, existing literature was surveyed. Data sufficient to set the dose was inadequate, and so an (n = 10) phase I-A calibration trial was performed, administering serine with and without other oral agents. We analyzed the trial and the literature data using pharmacokinetic (PK) modeling and statistical analysis. The therapeutic goal is to modulate specific serine-related metabolic pathways in the liver using the lowest possible dose which gives the desired effect since the upper bound was expected to be limited by toxicity. A standard PK approach, in which a common model structure was selected using a fit to data, yielded a model with a single central compartment corresponding to plasma, clearance from that compartment, and an endogenous source of serine. To improve conditioning, a parametric structure was changed to estimate ratios (bioavailability over volume, for example). Model fit quality was improved and the uncertainty in estimated parameters was reduced. Because of the particular interest in the fate of serine, the model was used to estimate whether serine is consumed in the gut, absorbed by the liver, or entered the blood in either a free state, or in a protein- or tissue-bound state that is not measured by our assay. The PK model structure was set up to represent relevant physiology, and this quantitative systems biology approach allowed a broader set of physiological data to be used to narrow parameter and prediction confidence intervals, and to better understand the biological meaning of the data. The model results allowed us to determine the optimal human dose for future trials, including a trial design component including IV and tracer studies. A key contribution is that we were able to use human physiological data from the literature to inform the PK model and to set reasonable bounds on parameters, and to improve model conditioning. Leveraging literature data produced a more predictive, useful model.