Computational intelligence is increasingly common in interactive systems in many domains, including health. Health coaching with conversational agents (CA) can reach wide populations, but the level of computational intelligence needed for a positive coaching experience is unclear. We conducted a study with sixteen individuals with diabetes and prediabetes who used a CA for health coaching, T2 Coach. Qualitative interviews revealed that participants saw T2 Coach as reliable in helping them stay on track with self-management, appreciated the flexibility in choosing personally meaningful goals and engaging on their own terms, and felt it provided encouragement and even compared it favorably with human coaches. However, they also noted that coaching experience could be improved with more fluid conversations, more tailoring to their personal preferences and lifestyles, and more sensitivity to specific contexts, all of which require more computational intelligence. We discuss implications and design directions for more intelligent coaching CA in health.

Transformers, and the attention mechanism in particular, have become ubiquitous in machine learning. Their success in modeling nonlocal, long-range correlations has led to their widespread adoption in natural language processing, computer vision, and time series problems. Neural operators, which map spaces of functions into spaces of functions, are necessarily both nonlinear and nonlocal if they are universal; it is thus natural to ask whether the attention mechanism can be used in the design of neural operators. Motivated by this, we study transformers in the function space setting. We formulate attention as a map between infinite dimensional function spaces and prove that the attention mechanism as implemented in practice is a Monte Carlo or finite difference approximation of this operator. The function space formulation allows for the design of transformer neural operators, a class of architectures designed to learn mappings between function spaces. In this paper, we state and prove the first universal approximation result for transformer neural operators, using only a slight modification of the architecture implemented in practice. The prohibitive cost of applying the attention operator to functions defined on multi-dimensional domains leads to the need for more efficient attention-based architectures. For this reason we also introduce a function space generalization of the patching strategy from computer vision, and introduce a class of associated neural operators. Numerical results, on an array of operator learning problems, demonstrate the promise of our approaches to function space formulations of attention and their use in neural operators.

Hybrid models composing mechanistic ODE-based dynamics with flexible and expressive neural network components have grown rapidly in popularity, especially in scientific domains where such ODE-based modeling offers important interpretability and validated causal grounding (e.g., for counterfactual reasoning). The incorporation of mechanistic models also provides inductive bias in standard blackbox modeling approaches, critical when learning from small datasets or partially observed, complex systems. Unfortunately, as the hybrid models become more flexible, the causal grounding provided by the mechanistic model can quickly be lost. We address this problem by leveraging another common source of domain knowledge: ranking of treatment effects for a set of interventions, even if the precise treatment effect is unknown. We encode this information in a causal loss that we combine with the standard predictive loss to arrive at a hybrid loss that biases our learning towards causally valid hybrid models. We demonstrate our ability to achieve a win-win, state-of-the-art predictive performance and causal validity, in the challenging task of modeling glucose dynamics post-exercise in individuals with type 1 diabetes.

OBJECTIVE: Prediction of physiological mechanics are important in medical practice because interventions are guided by predicted impacts of interventions. But prediction is difficult in medicine because medicine is complex and difficult to understand from data alone, and the data are sparse relative to the complexity of the generating processes. Computational methods can increase prediction accuracy, but prediction with clinical data is difficult because the data are sparse, noisy and nonstationary. This paper focuses on predicting physiological processes given sparse, non-stationary, electronic health record data in the intensive care unit using data assimilation (DA), a broad collection of methods that pair mechanistic models with inference methods. METHODS: A methodological pipeline embedding a glucose-insulin model into a new DA framework, the constrained ensemble Kalman filter (CEnKF) to forecast blood glucose was developed. The data include tube-fed patients whose nutrition, blood glucose, administered insulins and medications were extracted by hand due to their complexity and to ensure accuracy. The model was estimated using an individual's data as if they arrived in real-time, and the estimated model was run forward producing a forecast. Both constrained and unconstrained ensemble Kalman filters were estimated to compare the impact of constraints. Constraint boundaries, model parameter sets estimated, and data used to estimate the models were varied to investigate their influence on forecasting accuracy. Forecasting accuracy was evaluated according to mean squared error between the model-forecasted glucose and the measurements and by comparing distributions of measured glucose and forecast ensemble means. RESULTS: The novel CEnKF produced substantial gains in robustness and accuracy while minimizing the data requirements compared to the unconstrained ensemble Kalman filters. Administered insulin and tube-nutrition were important for accurate forecasting, but including glucose in IV medication delivery did not increase forecast accuracy. Model flexibility, controlled by constraint boundaries and estimated parameters, did influence forecasting accuracy. CONCLUSION: Accurate and robust physiological forecasting with sparse clinical data is possible with DA. Introducing constrained inference, particularly on unmeasured states and parameters, reduced forecast error and data requirements. The results are not particularly sensitive to model flexibility such as constraint boundaries, but over or under constraining increased forecasting errors.

Self-tracking can help personalize self-management interventions for chronic conditions like type 2 diabetes (T2D), but reflecting on personal data requires motivation and literacy. Machine learning (ML) methods can identify patterns, but a key challenge is making actionable suggestions based on personal health data. We introduce GlucoGoalie, which combines ML with an expert system to translate ML output into personalized nutrition goal suggestions for individuals with T2D. In a controlled experiment, participants with T2D found that goal suggestions were understandable and actionable. A 4-week in-the-wild deployment study showed that receiving goal suggestions augmented participants' self-discovery, choosing goals highlighted the multifaceted nature of personal preferences, and the experience of following goals demonstrated the importance of feedback and context. However, we identified tensions between abstract goals and concrete eating experiences and found static text too ambiguous for complex concepts. We discuss implications for ML-based interventions and the need for systems that offer more interactivity, feedback, and negotiation.

OBJECTIVE: To rapidly develop, validate, and implement a novel real-time mortality score for the COVID-19 pandemic that improves upon sequential organ failure assessment (SOFA) for decision support for a Crisis Standards of Care team. MATERIALS AND METHODS: We developed, verified, and deployed a stacked generalization model to predict mortality using data available in the electronic health record (EHR) by combining 5 previously validated scores and additional novel variables reported to be associated with COVID-19-specific mortality. We verified the model with prospectively collected data from 12 hospitals in Colorado between March 2020 and July 2020. We compared the area under the receiver operator curve (AUROC) for the new model to the SOFA score and the Charlson Comorbidity Index. RESULTS: The prospective cohort included 27 296 encounters, of which 1358 (5.0%) were positive for SARS-CoV-2, 4494 (16.5%) required intensive care unit care, 1480 (5.4%) required mechanical ventilation, and 717 (2.6%) ended in death. The Charlson Comorbidity Index and SOFA scores predicted mortality with an AUROC of 0.72 and 0.90, respectively. Our novel score predicted mortality with AUROC 0.94. In the subset of patients with COVID-19, the stacked model predicted mortality with AUROC 0.90, whereas SOFA had AUROC of 0.85. DISCUSSION: Stacked regression allows a flexible, updatable, live-implementable, ethically defensible predictive analytics tool for decision support that begins with validated models and includes only novel information that improves prediction. CONCLUSION: We developed and validated an accurate in-hospital mortality prediction score in a live EHR for automatic and continuous calculation using a novel model that improved upon SOFA.

[This corrects the article DOI: 10.1371/journal.pcbi.1005232.].

The increasing availability of health data and knowledge about computationally modeling human physiology opens new opportunities for personalized predictions in health. Yet little is known about how individuals interact and reason with personalized predictions. To explore these questions, we developed a smartphone app, GlucOracle, that uses self-tracking data of individuals with type 2 diabetes to generate personalized forecasts for post-meal blood glucose levels. We pilot-tested GlucOracle with two populations: members of an online diabetes community, knowledgeable about diabetes and technologically savvy; and individuals from a low socio-economic status community, characterized by high prevalence of diabetes, low literacy and limited experience with mobile apps. Individuals in both communities engaged with personal glucose forecasts and found them useful for adjusting immediate meal options, and planning future meals. However, the study raised new questions as to appropriate time, form, and focus of forecasts and suggested new research directions for personalized predictions in health.

Type 2 Diabetes Mellitus (T2DM) is a common chronic condition that requires management of one's lifestyle, including nutrition. Critically, patients often lack a clear understanding of how everyday meals impact their blood glucose. New predictive analytics approaches can provide personalized mealtime blood glucose forecasts. While communicating forecasts can be challenging, effective strategies for doing so remain little explored. In this study, we conducted focus groups with 13 participants to identify approaches to visualizing personalized blood glucose forecasts that can promote diabetes self-management and understand key styles and visual features that resonate with individuals with diabetes. Focus groups demonstrated that individuals rely on simple heuristics and tend to take a reactive approach to their health and nutrition management. Further, the study highlighted the need for simple and explicit, yet information-rich design. Effective visualizations were found to utilize common metaphors alongside words, numbers, and colors to convey a sense of authority and encourage action and learning.

We introduce data assimilation as a computational method that uses machine learning to combine data with human knowledge in the form of mechanistic models in order to forecast future states, to impute missing data from the past by smoothing, and to infer measurable and unmeasurable quantities that represent clinically and scientifically important phenotypes. We demonstrate the advantages it affords in the context of type 2 diabetes by showing how data assimilation can be used to forecast future glucose values, to impute previously missing glucose values, and to infer type 2 diabetes phenotypes. At the heart of data assimilation is the mechanistic model, here an endocrine model. Such models can vary in complexity, contain testable hypotheses about important mechanics that govern the system (eg, nutrition's effect on glucose), and, as such, constrain the model space, allowing for accurate estimation using very little data.

Objective: To develop and test a visual analytics tool to help clinicians identify systematic and clinically meaningful patterns in patient-generated data (PGD) while decreasing perceived information overload. Methods: Participatory design was used to develop Glucolyzer, an interactive tool featuring hierarchical clustering and a heatmap visualization to help registered dietitians (RDs) identify associative patterns between blood glucose levels and per-meal macronutrient composition for individuals with type 2 diabetes (T2DM). Ten RDs participated in a within-subjects experiment to compare Glucolyzer to a static logbook format. For each representation, participants had 25 minutes to examine 1 month of diabetes self-monitoring data captured by an individual with T2DM and identify clinically meaningful patterns. We compared the quality and accuracy of the observations generated using each representation. Results: Participants generated 50% more observations when using Glucolyzer (98) than when using the logbook format (64) without any loss in accuracy (69% accuracy vs 62%, respectively, p = .17). Participants identified more observations that included ingredients other than carbohydrates using Glucolyzer (36% vs 16%, p = .027). Fewer RDs reported feelings of information overload using Glucolyzer compared to the logbook format. Study participants displayed variable acceptance of hierarchical clustering. Conclusions: Visual analytics have the potential to mitigate provider concerns about the volume of self-monitoring data. Glucolyzer helped dietitians identify meaningful patterns in self-monitoring data without incurring perceived information overload. Future studies should assess whether similar tools can support clinicians in personalizing behavioral interventions that improve patient outcomes.

Objective: To study the effect on patient cohorts of mapping condition (diagnosis) codes from source billing vocabularies to a clinical vocabulary. Materials and Methods: Nine International Classification of Diseases, Ninth Revision, Clinical Modification (ICD9-CM) concept sets were extracted from eMERGE network phenotypes, translated to Systematized Nomenclature of Medicine - Clinical Terms concept sets, and applied to patient data that were mapped from source ICD9-CM and ICD10-CM codes to Systematized Nomenclature of Medicine - Clinical Terms codes using Observational Health Data Sciences and Informatics (OHDSI) Observational Medical Outcomes Partnership (OMOP) vocabulary mappings. The original ICD9-CM concept set and a concept set extended to ICD10-CM were used to create patient cohorts that served as gold standards. Results: Four phenotype concept sets were able to be translated to Systematized Nomenclature of Medicine - Clinical Terms without ambiguities and were able to perform perfectly with respect to the gold standards. The other 5 lost performance when 2 or more ICD9-CM or ICD10-CM codes mapped to the same Systematized Nomenclature of Medicine - Clinical Terms code. The patient cohorts had a total error (false positive and false negative) of up to 0.15% compared to querying ICD9-CM source data and up to 0.26% compared to querying ICD9-CM and ICD10-CM data. Knowledge engineering was required to produce that performance; simple automated methods to generate concept sets had errors up to 10% (one outlier at 250%). Discussion: The translation of data from source vocabularies to Systematized Nomenclature of Medicine - Clinical Terms (SNOMED CT) resulted in very small error rates that were an order of magnitude smaller than other error sources. Conclusion: It appears possible to map diagnoses from disparate vocabularies to a single clinical vocabulary and carry out research using a single set of definitions, thus improving efficiency and transportability of research.

Type 2 diabetes leads to premature death and reduced quality of life for 8% of Americans. Nutrition management is critical to maintaining glycemic control, yet it is difficult to achieve due to the high individual differences in glycemic response to nutrition. Anticipating glycemic impact of different meals can be challenging not only for individuals with diabetes, but also for expert diabetes educators. Personalized computational models that can accurately forecast an impact of a given meal on an individual's blood glucose levels can serve as the engine for a new generation of decision support tools for individuals with diabetes. However, to be useful in practice, these computational engines need to generate accurate forecasts based on limited datasets consistent with typical self-monitoring practices of individuals with type 2 diabetes. This paper uses three forecasting machines: (i) data assimilation, a technique borrowed from atmospheric physics and engineering that uses Bayesian modeling to infuse data with human knowledge represented in a mechanistic model, to generate real-time, personalized, adaptable glucose forecasts; (ii) model averaging of data assimilation output; and (iii) dynamical Gaussian process model regression. The proposed data assimilation machine, the primary focus of the paper, uses a modified dual unscented Kalman filter to estimate states and parameters, personalizing the mechanistic models. Model selection is used to make a personalized model selection for the individual and their measurement characteristics. The data assimilation forecasts are empirically evaluated against actual postprandial glucose measurements captured by individuals with type 2 diabetes, and against predictions generated by experienced diabetes educators after reviewing a set of historical nutritional records and glucose measurements for the same individual. The evaluation suggests that the data assimilation forecasts compare well with specific glucose measurements and match or exceed in accuracy expert forecasts. We conclude by examining ways to present predictions as forecast-derived range quantities and evaluate the comparative advantages of these ranges.

OBJECTIVE: To investigate how individuals with diabetes and diabetes educators reason about data collected through self-monitoring and to draw implications for the design of data-driven self-management technologies. MATERIALS AND METHODS: Ten individuals with diabetes (six type 1 and four type 2) and 2 experienced diabetes educators were presented with a set of self-monitoring data captured by an individual with type 2 diabetes. The set included digital images of meals and their textual descriptions, and blood glucose (BG) readings captured before and after these meals. The participants were asked to review a set of meals and associated BG readings, explain differences in postprandial BG levels for these meals, and predict postprandial BG levels for the same individual for a different set of meals. Researchers compared conclusions and predictions reached by the participants with those arrived at by quantitative analysis of the collected data. RESULTS: The participants used both macronutrient composition of meals, most notably the inclusion of carbohydrates, and names of dishes and ingredients to reason about changes in postprandial BG levels. Both individuals with diabetes and diabetes educators reported difficulties in generating predictions of postprandial BG; their predictions varied in their correlations with the actual captured readings from r = 0.008 to r = 0.75. CONCLUSION: Overall, the study showed that identifying trends in the data collected with self-monitoring is a complex process, and that conclusions reached by both individuals with diabetes and diabetes educators are not always reliable. This suggests the need for new ways to facilitate individuals' reasoning with informatics interventions.