Transforming Big Data into AI-ready data for nutrition and obesity research.

OBJECTIVE: Big Data are increasingly used in obesity and nutrition research to gain new insights and derive personalized guidance; however, this data in raw form are often not usable. Substantial preprocessing, which requires machine learning (ML), human judgment, and specialized software, is required to transform Big Data into artificial intelligence (AI)- and ML-ready data. These preprocessing steps are the most complex part of the entire modeling pipeline. Understanding the complexity of these steps by the end user is critical for reducing misunderstanding, faulty interpretation, and erroneous downstream conclusions. METHODS: We reviewed three popular obesity/nutrition Big Data sources: microbiome, metabolomics, and accelerometry. The preprocessing pipelines, specialized software, challenges, and how decisions impact final AI- and ML-ready products were detailed. RESULTS: Opportunities for advances to improve quality control, speed of preprocessing, and intelligent end user consumption were presented. CONCLUSIONS: Big Data have the exciting potential for identifying new modifiable factors that impact obesity research. However, to ensure accurate interpretation of conclusions arising from Big Data, the choices involved in preparing AI- and ML-ready data need to be transparent to investigators and clinicians relying on the conclusions.

Objective Determination of Eating Occasion Timing: Combining Self-Report, Wrist Motion, and Continuous Glucose Monitoring to Detect Eating Occasions in Adults With Prediabetes and Obesity.

BACKGROUND: Accurately identifying eating patterns, specifically the timing, frequency, and distribution of eating occasions (EOs), is important for assessing eating behaviors, especially for preventing and managing obesity and type 2 diabetes (T2D). However, existing methods to study EOs rely on self-report, which may be prone to misreporting and bias and has a high user burden. Therefore, objective methods are needed. METHODS: We aim to compare EO timing using objective and subjective methods. Participants self-reported EO with a smartphone app (self-report [SR]), wore the ActiGraph GT9X on their dominant wrist, and wore a continuous glucose monitor (CGM, Abbott Libre Pro) for 10 days. EOs were detected from wrist motion (WM) using a motion-based classifier and from CGM using a simulation-based system. We described EO timing and explored how timing identified with WM and CGM compares with SR. RESULTS: Participants (n = 39) were 59 ± 11 years old, mostly female (62%) and White (51%) with a body mass index (BMI) of 34.2 ± 4.7 kg/m2. All had prediabetes or moderately controlled T2D. The median time-of-day first EO (and interquartile range) for SR, WM, and CGM were 08:24 (07:00-09:59), 9:42 (07:46-12:26), and 06:55 (04:23-10:03), respectively. The median last EO for SR, WM, and CGM were 20:20 (16:50-21:42), 20:12 (18:30-21:41), and 21:43 (20:35-22:16), respectively. The overlap between SR and CGM was 55% to 80% of EO detected with tolerance periods of ±30, 60, and 120 minutes. The overlap between SR and WM was 52% to 65% EO detected with tolerance periods of ±30, 60, and 120 minutes. CONCLUSION: The continuous glucose monitor and WM detected overlapping but not identical meals and may provide complementary information to self-reported EO.

Less is more: information needs, information wants, and what makes causal models useful.

Each day people make decisions about complex topics such as health and personal finances. Causal models of these domains have been created to aid decisions, but the resulting models are often complex and it is not known whether people can use them successfully. We investigate the trade-off between simplicity and complexity in decision making, testing diagrams tailored to target choices (Experiments 1 and 2), and with relevant causal paths highlighted (Experiment 3), finding that simplicity or directing attention to simple causal paths leads to better decisions. We test the boundaries of this effect (Experiment 4), finding that including a small amount of information beyond that related to the target answer has a detrimental effect. Finally, we examine whether people know what information they need (Experiment 5). We find that simple, targeted, information still leads to the best decisions, while participants who believe they do not need information or seek out the most complex information performed worse.

Simulating Realistic Continuous Glucose Monitor Time Series By Data Augmentation.

BACKGROUND: Simulated data are a powerful tool for research, enabling benchmarking of blood glucose (BG) forecasting and control algorithms. However, expert created models provide an unrealistic view of real-world performance, as they lack the features that make real data challenging, while black-box approaches such as generative adversarial networks do not enable systematic tests to diagnose model performance. METHODS: To address this, we propose a method that learns missingness and error properties of continuous glucose monitor (CGM) data collected from people with type 1 diabetes (OpenAPS, OhioT1DM, RCT, and Racial-Disparity), and then augments simulated BG data with these properties. On the task of BG forecasting, we test how well our method brings performance closer to that of real CGM data compared with current simulation practices for missing data (random dropout) and error (Gaussian noise, CGM error model). RESULTS: Our methods had the smallest performance difference versus real data compared with random dropout and Gaussian noise when individually testing the effects of missing data and error on simulated BG in most cases. When combined, our approach was significantly better than Gaussian noise and random dropout for all data sets except OhioT1DM. Our error model significantly improved results on diverse data sets. CONCLUSIONS: We find a significant gap between BG forecasting performance on simulated and real data, and our method can be used to close this gap. This will enable researchers to rigorously test algorithms and provide realistic estimates of real-world performance without overfitting to real data or at the expense of data collection.

Classification of Level of Consciousness in a Neurological ICU Using Physiological Data.

BACKGROUND: Impaired consciousness is common in intensive care unit (ICU) patients, and an individual's degree of consciousness is crucial to determining their care and prognosis. However, there are no methods that continuously monitor consciousness and alert clinicians to changes. We investigated the use of physiological signals collected in the ICU to classify levels of consciousness in critically ill patients. METHODS: We studied 61 patients with subarachnoid hemorrhage (SAH) and 178 patients with intracerebral hemorrhage (ICH) from the neurological ICU at Columbia University Medical Center in a retrospective observational study of prospectively collected data. The level of consciousness was determined on the basis of neurological examination and mapped to comatose, vegetative state or unresponsive wakefulness syndrome (VS/UWS), minimally conscious minus state (MCS-), and command following. For each physiological signal, we extracted time-series features and performed classification using extreme gradient boosting on multiple clinically relevant tasks across subsets of physiological signals. We applied this approach independently on both SAH and ICH patient groups for three sets of variables: (1) a minimal set common to most hospital patients (e.g., heart rate), (2) variables available in most ICUs (e.g., body temperature), and (3) an extended set recorded mainly in neurological ICUs (absent for the ICH patient group; e.g., brain temperature). RESULTS: On the commonly performed classification task of VS/UWS versus MCS-, we achieved an area under the receiver operating characteristic curve (AUROC) in the SAH patient group of 0.72 (sensitivity 82%, specificity 57%; 95% confidence interval [CI] 0.63-0.81) using the extended set, 0.69 (sensitivity 83%, specificity 51%; 95% CI 0.59-0.78) on the variable set available in most ICUs, and 0.69 (sensitivity 56%, specificity 78%; 95% CI 0.60-0.78) on the minimal set. In the ICH patient group, AUROC was 0.64 (sensitivity 56%, specificity 65%; 95% CI 0.55-0.74) using the minimal set and 0.61 (sensitivity 50%, specificity 80%; 95% CI 0.51-0.71) using the variables available in most ICUs. CONCLUSIONS: We find that physiological signals can be used to classify states of consciousness for patients in the ICU. Building on this with intraday assessments and increasing sensitivity and specificity may enable alarm systems that alert physicians to changes in consciousness and frequent monitoring of consciousness throughout the day, both of which may improve patient care and outcomes.

Artificial Intelligence for Predicting and Diagnosing Complications of Diabetes.

Artificial intelligence can use real-world data to create models capable of making predictions and medical diagnosis for diabetes and its complications. The aim of this commentary article is to provide a general perspective and present recent advances on how artificial intelligence can be applied to improve the prediction and diagnosis of six significant complications of diabetes including (1) gestational diabetes, (2) hypoglycemia in the hospital, (3) diabetic retinopathy, (4) diabetic foot ulcers, (5) diabetic peripheral neuropathy, and (6) diabetic nephropathy.

Machine learning modeling practices to support the principles of AI and ethics in nutrition research.

BACKGROUND: Nutrition research is relying more on artificial intelligence and machine learning models to understand, diagnose, predict, and explain data. While artificial intelligence and machine learning models provide powerful modeling tools, failure to use careful and well-thought-out modeling processes can lead to misleading conclusions and concerns surrounding ethics and bias. METHODS: Based on our experience as reviewers and journal editors in nutrition and obesity, we identified the most frequently omitted best practices from statistical modeling and how these same practices extend to machine learning models. We next addressed areas required for implementation of machine learning that are not included in commercial software packages. RESULTS: Here, we provide a tutorial on best artificial intelligence and machine learning modeling practices that can reduce potential ethical problems with a checklist and guiding principles to aid nutrition researchers in developing, evaluating, and implementing artificial intelligence and machine learning models in nutrition research. CONCLUSION: The quality of AI/ML modeling in nutrition research requires iterative and tailored processes to mitigate against potential ethical problems or to predict conclusions that are free of bias.

Health Information Sourcing and Health Knowledge Quality: Repeated Cross-sectional Survey.

BACKGROUND: People's health-related knowledge influences health outcomes, as this knowledge may influence whether individuals follow advice from their doctors or public health agencies. Yet, little attention has been paid to where people obtain health information and how these information sources relate to the quality of knowledge. OBJECTIVE: We aim to discover what information sources people use to learn about health conditions, how these sources relate to the quality of their health knowledge, and how both the number of information sources and health knowledge change over time. METHODS: We surveyed 200 different individuals at 12 time points from March through September 2020. At each time point, we elicited participants' knowledge about causes, risk factors, and preventative interventions for 8 viral (Ebola, common cold, COVID-19, Zika) and nonviral (food allergies, amyotrophic lateral sclerosis [ALS], strep throat, stroke) illnesses. Participants were further asked how they learned about each illness and to rate how much they trust various sources of health information. RESULTS: We found that participants used different information sources to obtain health information about common illnesses (food allergies, strep throat, stroke) compared to emerging illnesses (Ebola, common cold, COVID-19, Zika). Participants relied mainly on news media, government agencies, and social media for information about emerging illnesses, while learning about common illnesses from family, friends, and medical professionals. Participants relied on social media for information about COVID-19, with their knowledge accuracy of COVID-19 declining over the course of the pandemic. The number of information sources participants used was positively correlated with health knowledge quality, though there was no relationship with the specific source types consulted. CONCLUSIONS: Building on prior work on health information seeking and factors affecting health knowledge, we now find that people systematically consult different types of information sources by illness type and that the number of information sources people use affects the quality of individuals' health knowledge. Interventions to disseminate health information may need to be targeted to where individuals are likely to seek out information, and these information sources differ systematically by illness type.

Using Momentary Assessment and Machine Learning to Identify Barriers to Self-management in Type 1 Diabetes: Observational Study.

BACKGROUND: For adolescents living with type 1 diabetes (T1D), completion of multiple daily self-management tasks, such as monitoring blood glucose and administering insulin, can be challenging because of psychosocial and contextual barriers. These barriers are hard to assess accurately and specifically by using traditional retrospective recall. Ecological momentary assessment (EMA) uses mobile technologies to assess the contexts, subjective experiences, and psychosocial processes that surround self-management decision-making in daily life. However, the rich data generated via EMA have not been frequently examined in T1D or integrated with machine learning analytic approaches. OBJECTIVE: The goal of this study is to develop a machine learning algorithm to predict the risk of missed self-management in young adults with T1D. To achieve this goal, we train and compare a number of machine learning models through a learned filtering architecture to explore the extent to which EMA data were associated with the completion of two self-management behaviors: mealtime self-monitoring of blood glucose (SMBG) and insulin administration. METHODS: We analyzed data from a randomized controlled pilot study using machine learning-based filtering architecture to investigate whether novel information related to contextual, psychosocial, and time-related factors (ie, time of day) relate to self-management. We combined EMA-collected contextual and insulin variables via the MyDay mobile app with Bluetooth blood glucose data to construct machine learning classifiers that predicted the 2 self-management behaviors of interest. RESULTS: With 1231 day-level SMBG frequency counts for 45 participants, demographic variables and time-related variables were able to predict whether daily SMBG was below the clinical threshold of 4 times a day. Using the 1869 data points derived from app-based EMA data of 31 participants, our learned filtering architecture method was able to infer nonadherence events with high accuracy and precision. Although the recall score is low, there is high confidence that the nonadherence events identified by the model are truly nonadherent. CONCLUSIONS: Combining EMA data with machine learning methods showed promise in the relationship with risk for nonadherence. The next steps include collecting larger data sets that would more effectively power a classifier that can be deployed to infer individual behavior. Improvements in individual self-management insights, behavioral risk predictions, enhanced clinical decision-making, and just-in-time patient support in diabetes could result from this type of approach.

Hierarchical Information Criterion for Variable Abstraction.

Large biomedical datasets can contain thousands of variables, creating challenges for machine learning tasks such as causal inference and prediction. Feature selection and ranking methods have been developed to reduce the number of variables and determine which are most important. However in many cases, such as in classification from diagnosis codes, ontologies, and controlled vocabularies, we must choose not only which variables to include but also at what level of granularity. ICD-9 codes, for example, are arranged in a hierarchy, and a user must decide at what level codes should be analyzed. Thus it is currently up to a researcher to decide whether to use any diagnosis of diabetes or whether to distinguish between specific forms, such as Type 2 diabetes with renal complications versus without mention of complications. Currently, there is no existing method that can automatically make this determination and methods for feature selection do not exploit this hierarchical information, which is found in other areas including nutrition (hierarchies of foods), and bioinformatics (hierarchical relationship of genes). To address this, we propose a novel Hierarchical Information Criterion (HIC) that builds on mutual information and allows fully automated abstraction of variables. Using HIC allows us to rank hierarchical features and select the ones with the highest score. We show that this significantly improves performance by an average AUROC of 0.053 over traditional feature selection methods and hand crafted features on two mortality prediction tasks using MIMIC-III ICU data. Our method also improves on the state of the art (Fu et al., 2019) with an AUROC increase from 0.819 to 0.887.

How causal information affects decisions.

BACKGROUND: Causality is inherently linked to decision-making, as causes let us better predict the future and intervene to change it by showing which variables have the capacity to affect others. Recent advances in machine learning have made it possible to learn causal models from observational data. While these models have the potential to aid human decisions, it is not yet known whether the output of these algorithms improves decision-making. That is, causal inference methods have been evaluated on their accuracy at uncovering ground truth, but not the utility of such output for human consumption. Simply presenting more information to people may not have the intended effects, particularly when they must combine this information with their existing knowledge and beliefs. While psychological studies have shown that causal models can be used to choose interventions and predict outcomes, that work has not tested structures of the complexity found in machine learning, or how such information is interpreted in the context of existing knowledge. RESULTS: Through experiments on Amazon Mechanical Turk, we study how people use causal information to make everyday decisions about diet, health, and personal finance. Our first experiment, using decisions about maintaining bodyweight, shows that causal information can actually lead to worse decisions than no information at all. In Experiment 2, we test decisions about diabetes management, where some participants have personal domain experience and others do not. We find that individuals without such experience are aided by causal information, while individuals with experience do worse. Finally, our last two experiments probe how prior experience interacts with causal information. We find that while causal information reduces confidence in individuals with prior experience, it has the opposite effect on those without experience. In Experiment 4 we show that our results are not due to an inability to use causal models, and that they may be due to familiarity with a domain rather than actual knowledge. CONCLUSION: While causal inference can potentially lead to more informed decisions, we find that more work is needed to make causal models useful for the types of decisions found in daily life.

Comparing Machine Learning Techniques for Blood Glucose Forecasting Using Free-living and Patient Generated Data.

Managing a chronic disease like Type 1 diabetes (T1D) is both challenging and time consuming, but new technologies that allow continuous measurement of glucose and delivery of insulin have led to significant improvements. The development of an artificial pancreas (AP), which algorithmically determines insulin dosing and delivers insulin in a fully automated way, may transform T1D care but it is not yet widely available. Patient-led alternatives, like the Open Artificial Pancreas (OpenAPS), are being used by hundreds of individuals and have also led to a dramatic increase in the availability of patient generated health data (PGHD). All APs require an accurate forecast of blood glucose (BG). While there have been efforts to develop better forecasts and apply new ML techniques like deep learning to this problem, methods are often tested on small controlled datasets that do not indicate how they may perform in reality - and the most advanced methods have not always outperformed the simplest. We introduce a rigorous comparison of BG forecasting using both a small controlled research dataset and large heterogeneous PGHD. Our comparison advances the state of the art in BG forecasting by providing insight into how methods may fare when moving beyond small controlled studies to real-world use.

Automated meal detection from continuous glucose monitor data through simulation and explanation.

BACKGROUND: Artificial pancreas systems aim to reduce the burden of type 1 diabetes by automating insulin dosing. These systems link a continuous glucose monitor (CGM) and insulin pump with a control algorithm, but require users to announce meals, without which the system can only react to the rise in blood glucose. OBJECTIVE: We investigate whether CGM data can be used to automatically infer meals in daily life even in the presence of physical activity, which can raise or lower blood glucose. MATERIALS AND METHODS: We propose a novel meal detection algorithm that combines simulations with CGM, insulin pump, and heart rate monitor data. When observed and predicted glucose differ, our algorithm uses simulations to test whether a meal may explain this difference. We evaluated our method on simulated data and real-world data from individuals with type 1 diabetes. RESULTS: In simulated data, we detected meals earlier and with higher accuracy than was found in prior work (25.7 minutes, 1.2 g error; compared with 48.3 minutes, 17.2 g error). In real-world data, we discovered a larger number of plausible meals than was found in prior work (30 meals, 76.7% accepted; compared with 33 meals, 39.4% accepted). DISCUSSION: Prior research attempted meal detection from CGM, but had delays and lower accuracy in real data or did not allow for physical activity. Our approach can be used to improve insulin dosing in an artificial pancreas and trigger reminders for missed meal boluses. CONCLUSIONS: We demonstrate that meal information can be robustly inferred from CGM and body-worn sensor data, even in challenging environments of daily life.

Lagged Correlations among Physiological Variables as Indicators of Consciousness in Stroke Patients.

Consciousness is a highly significant indicator of an ICU patient's condition but there is still no method to automatically measure it. Instead, time consuming and subjective assessments are used. However, many brain and physiologic variables are measured continuously in neurological ICU, and could be used as indicators for consciousness. Since many biological variables are highly correlated to maintain homeostasis, we examine whether changes in time lags between correlated variables may relate to changes in consciousness. We introduce new methods to identify changes in the time lag of correlations, which better handle noisy multimodal physiological data and fluctuating lags. On neurological ICU data from subarachnoid hemorrhage patients, we find that correlations among variables related to brain physiology or respiration have significantly longer lags inpatients with decreased levels of consciousness than in patients with higher levels of consciousness. This suggests that physiological data could potentially be used to automatically assess consciousness.

Using Domain Knowledge to Overcome Latent Variables in Causal Inference from Time Series.

Increasingly large observational datasets from healthcare and social media may allow new types of causal inference. However, these data are often missing key variables, increasing the chance of finding spurious causal relationships due to confounding. While methods exist for causal inference with latent variables in static cases, temporal relationships are more challenging, as varying time lags make latent causes more difficult to uncover and approaches often have significantly higher computational complexity. To address this, we make the key observation that while a variable may be latent in one dataset, it may be observed in another, or we may have domain knowledge about its effects. We propose a computationally efficient method that overcomes latent variables by using prior knowledge to reconstruct data for unobserved variables, while remaining robust to cases when the knowledge is wrong or does not apply. On simulated data, our approach outperforms the state of the art with a lower false discovery rate for causal inference. On real-world data from individuals with Type 1 diabetes, we show that our approach can discover causal relationships involving unmeasured meals and exercise.

Automated Estimation of Food Type from Body-worn Audio and Motion Sensors in Free-Living Environments.

Nutrition is fundamental to maintaining health, managing chronic diseases, and preventing illness, but unlike physical activity there is not yet a way to unobtrusively and automatically measure nutrition. While recent work has shown that body-worn sensors can be used to identify meal times, to have an impact on health and fully replace manual food logs, we need to identify not only when someone is eating, but what they are consuming. However, it is challenging to collect labeled data in daily life, while lab data does not always generalize to reality. To address this, we develop new algorithms for semi-supervised hierarchical classification that enable higher accuracy when training on data with weak labels. Using this approach, we present the first results on automated classification of foods consumed in data collected from body-worn audio and motion sensors in free-living environments. We show that by exploiting a mix of lab and free-living data, we can achieve a classification accuracy of 88% on unrestricted meals (e.g. stir fry, pizza, salad) in unrestricted environments such as home and restaurants. Ultimately, this lays the foundation for body-worn devices that can calculate calories and macronutrients by identifying food type and quantity.

Automated Identification of Causal Moderators in Time-Series Data.

Causal inference is often taken to mean finding links between individual variables. However in many real-world cases, such as in biological systems, relationships are more complex, with groups of factors needed to produce an effect, or some factors only modifying other relationships rather than producing outcomes alone. For instance, weight may alter the efficacy of a drug without causing side effects itself. Such moderating factors may change the timing, intensity, or probability of a causal relationship. Distinguishing moderators from genuine causes can lead to more effective medical interventions, and better strategies for bringing about a desired effect, since a moderator alone is ineffective. However, there have not yet been algorithms to automatically infer moderators in a large-scale automated way, and they cannot be easily read off from causal graphs. We introduce a set of temporal logic rules to automatically identify the asymmetric roles of causes and moderators in a computationally efficient manner. Experiments on simulated data demonstrate that even in challenging cases we can find moderators and avoid confounding, and on real neurological ICU data we show how the approach can find more descriptive and meaningful relationships than the state of the art.

Replicability, Reproducibility, and Agent-based Simulation of Interventions.

Secondary use of medical data and use of observational data for causal inference has been growing. Yet these data bring many challenges such as confounding due to unobserved variables and variation in medical processes across settings. Further, while methods exist to handle some of these problems, researchers lack ground truth to evaluate these methods. When a finding is not replicated across multiple sites, it is unknown whether this is a failure of an algorithm, a genuine difference between populations, or an artifact of structural differences between the sites. We show how agent-based simulation of medical interventions can be used to explore how bias, error, and variation across settings affect inference. Our approach enables users to model not only interventions and outcomes, but also the complex interaction between patients with different risks of mortality and providers with different observed and latent treatment effects. Ultimately we propose that such simulations can be used to better evaluate the behavior of new methods with known ground truth and better calculate sample size for EHR-based studies.

Using uncertain data from body-worn sensors to gain insight into type 1 diabetes.

The amount of observational data available for research is growing rapidly with the rise of electronic health records and patient-generated data. However, these data bring new challenges, as data collected outside controlled environments and generated for purposes other than research may be error-prone, biased, or systematically missing. Analysis of these data requires methods that are robust to such challenges, yet methods for causal inference currently only handle uncertainty at the level of causal relationships - rather than variables or specific observations. In contrast, we develop a new approach for causal inference from time series data that allows uncertainty at the level of individual data points, so that inferences depend more strongly on variables and individual observations that are more certain. In the limit, a completely uncertain variable will be treated as if it were not measured. Using simulated data we demonstrate that the approach is more accurate than the state of the art, making substantially fewer false discoveries. Finally, we apply the method to a unique set of data collected from 17 individuals with type 1 diabetes mellitus (T1DM) in free-living conditions over 72h where glucose levels, insulin dosing, physical activity and sleep are measured using body-worn sensors. These data often have high rates of error that vary across time, but we are able to uncover the relationships such as that between anaerobic activity and hyperglycemia. Ultimately, better modeling of uncertainty may enable better translation of methods to free-living conditions, as well as better use of noisy and uncertain EHR data.

Causal Structure of Brain Physiology after Brain Injury from Subarachnoid Hemorrhage.

High frequency physiologic data are routinely generated for intensive care patients. While massive amounts of data make it difficult for clinicians to extract meaningful signals, these data could provide insight into the state of critically ill patients and guide interventions. We develop uniquely customized computational methods to uncover the causal structure within systemic and brain physiologic measures recorded in a neurological intensive care unit after subarachnoid hemorrhage. While the data have many missing values, poor signal-to-noise ratio, and are composed from a heterogeneous patient population, our advanced imputation and causal inference techniques enable physiologic models to be learned for individuals. Our analyses confirm that complex physiologic relationships including demand and supply of oxygen underlie brain oxygen measurements and that mechanisms for brain swelling early after injury may differ from those that develop in a delayed fashion. These inference methods will enable wider use of ICU data to understand patient physiology.