Population physiology: leveraging electronic health record data to understand human endocrine dynamics

Interpretable physiological forecasting in the ICU using constrained data assimilation and electronic health record data

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.

A simple modeling framework for prediction in the human glucose-insulin system

Forecasting blood glucose (BG) levels with routinely collected data is useful for glycemic management. BG dynamics are nonlinear, complex, and nonstationary, which can be represented by nonlinear models. However, the sparsity of routinely collected data creates parameter identifiability issues when high-fidelity complex models are used, thereby resulting in inaccurate forecasts. One can use models with reduced physiological fidelity for robust and accurate parameter estimation and forecasting with sparse data. For this purpose, we approximate the nonlinear dynamics of BG regulation by a linear stochastic differential equation: we develop a linear stochastic model, which can be specialized to different settings: type 2 diabetes mellitus (T2DM) and intensive care unit (ICU), with different choices of appropriate model functions. The model includes deterministic terms quantifying glucose removal from the bloodstream through the glycemic regulation system and representing the effect of nutrition and externally delivered insulin. The stochastic term encapsulates the BG oscillations. The model output is in the form of an expected value accompanied by a band around this value. The model parameters are estimated patient-specifically, leading to personalized models. The forecasts consist of values for BG mean and variation, quantifying possible high and low BG levels. Such predictions have potential use for glycemic management as part of control systems. We present experimental results on parameter estimation and forecasting in T2DM and ICU settings. We compare the model's predictive capability with two different nonlinear models built for T2DM and ICU contexts to have a sense of the level of prediction achieved by this model.

Incorporating real-world evidence into the development of patient blood glucose prediction algorithms for the ICU

Objective

Glycemic control is an important component of critical care. We present a data-driven method for predicting intensive care unit (ICU) patient response to glycemic control protocols while accounting for patient heterogeneity and variations in care.

Materials and Methods

Using electronic medical records (EMRs) of 18 961 ICU admissions from the MIMIC-III dataset, including 318 574 blood glucose measurements, we train and validate a gradient boosted tree machine learning (ML) algorithm to forecast patient blood glucose and a 95% prediction interval at 2-hour intervals. The model uses as inputs irregular multivariate time series data relating to recent in-patient medical history and glycemic control, including previous blood glucose, nutrition, and insulin dosing.

Results

Our forecasting model using routinely collected EMRs achieves performance comparable to previous models developed in planned research studies using continuous blood glucose monitoring. Model error, expressed as mean absolute percentage error is 16.5%–16.8%, with Clarke error grid analysis demonstrating that 97% of predictions would be clinically acceptable. The 95% prediction intervals achieve near intended coverage at 93%–94%.

Discussion

ML algorithms built on observational data sources, such as EMRs, present a promising approach for personalization and automation of glycemic control in critical care. Future research may benefit from applying a combination of methodologies and data sources to develop robust methodologies that account for the variations seen in ICU patients and difficultly in detecting the extremes of observed blood glucose values.

Conclusion

We demonstrate that EMRs can be used to train ML algorithms that may be suitable for incorporation into ICU decision support systems.

Delay-induced uncertainty for a paradigmatic glucose-insulin model

Medical practice in the intensive care unit is based on the assumption that physiological systems such as the human glucose-insulin system are predictable. We demonstrate that delay within the glucose-insulin system can induce sustained temporal chaos, rendering the system unpredictable. Specifically, we exhibit such chaos for the ultradian glucose-insulin model. This well-validated, finite-dimensional model represents feedback delay as a three-stage filter. Using the theory of rank one maps from smooth dynamical systems, we precisely explain the nature of the resulting delay-induced uncertainty (DIU). We develop a framework one may use to diagnose DIU in a general oscillatory dynamical system. For infinite-dimensional delay systems, no analog of the theory of rank one maps exists. Nevertheless, we show that the geometric principles encoded in our DIU framework apply to such systems by exhibiting sustained temporal chaos for a linear shear flow. Our results are potentially broadly applicable because delay is ubiquitous throughout mathematical physiology.

The parameter Houlihan: A solution to high-throughput identifiability indeterminacy for brutally ill-posed problems

One way to interject knowledge into clinically impactful forecasting is to use data assimilation, a nonlinear regression that projects data onto a mechanistic physiologic model, instead of a set of functions, such as neural networks. Such regressions have an advantage of being useful with particularly sparse, non-stationary clinical data. However, physiological models are often nonlinear and can have many parameters, leading to potential problems with parameter identifiability, or the ability to find a unique set of parameters that minimize forecasting error. The identifiability problems can be minimized or eliminated by reducing the number of parameters estimated, but reducing the number of estimated parameters also reduces the flexibility of the model and hence increases forecasting error. We propose a method, the parameter Houlihan, that combines traditional machine learning techniques with data assimilation, to select the right set of model parameters to minimize forecasting error while reducing identifiability problems. The method worked well: the data assimilation-based glucose forecasts and estimates for our cohort using the Houlihan-selected parameter sets generally also minimize forecasting errors compared to other parameter selection methods such as by-hand parameter selection. Nevertheless, the forecast with the lowest forecast error does not always accurately represent physiology, but further advancements of the algorithm provide a path for improving physiologic fidelity as well. Our hope is that this methodology represents a first step toward combining machine learning with data assimilation and provides a lower-threshold entry point for using data assimilation with clinical data by helping select the right parameters to estimate.

Mechanistic machine learning: how data assimilation leverages physiologic knowledge using Bayesian inference to forecast the future, infer the present, and phenotype

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.

When to re-order laboratory tests? Learning laboratory test shelf-life

Most laboratory results are valid for only a certain time period (laboratory tests shelf-life), after which they are outdated and the test needs to be re-administered. Currently, laboratory test shelf-lives are not centrally available anywhere but the implicit knowledge of doctors. In this work we propose an automated method to learn laboratory test-specific shelf-life by identifying prevalent laboratory test order patterns in electronic health records. The resulting shelf-lives performed well in the evaluation of internal validity, clinical interpretability, and external validity.

Methodological variations in lagged regression for detecting physiologic drug effects in EHR data

We studied how lagged linear regression can be used to detect the physiologic effects of drugs from data in the electronic health record (EHR). We systematically examined the effect of methodological variations ((i) time series construction, (ii) temporal parameterization, (iii) intra-subject normalization, (iv) differencing (lagged rates of change achieved by taking differences between consecutive measurements), (v) explanatory variables, and (vi) regression models) on performance of lagged linear methods in this context. We generated two gold standards (one knowledge-base derived, one expert-curated) for expected pairwise relationships between 7 drugs and 4 labs, and evaluated how the 64 unique combinations of methodological perturbations reproduce the gold standards. Our 28 cohorts included patients in the Columbia University Medical Center/NewYork-Presbyterian Hospital clinical database, and ranged from 2820 to 79,514 patients with between 8 and 209 average time points per patient. The most accurate methods achieved AUROC of 0.794 for knowledge-base derived gold standard (95%CI [0.741, 0.847]) and 0.705 for expert-curated gold standard (95% CI [0.629, 0.781]). We observed a mean AUROC of 0.633 (95%CI [0.610, 0.657], expert-curated gold standard) across all methods that re-parameterize time according to sequence and use either a joint autoregressive model with time-series differencing or an independent lag model without differencing. The complement of this set of methods achieved a mean AUROC close to 0.5, indicating the importance of these choices. We conclude that time-series analysis of EHR data will likely rely on some of the beneficial pre-processing and modeling methodologies identified, and will certainly benefit from continued careful analysis of methodological perturbations. This study found that methodological variations, such as pre-processing and representations, have a large effect on results, exposing the importance of thoroughly evaluating these components when comparing machine-learning methods.

High-fidelity phenotyping: richness and freedom from bias

Electronic health record phenotyping is the use of raw electronic health record data to assert characterizations about patients. Researchers have been doing it since the beginning of biomedical informatics, under different names. Phenotyping will benefit from an increasing focus on fidelity, both in the sense of increasing richness, such as measured levels, degree or severity, timing, probability, or conceptual relationships, and in the sense of reducing bias. Research agendas should shift from merely improving binary assignment to studying and improving richer representations. The field is actively researching new temporal directions and abstract representations, including deep learning. The field would benefit from research in nonlinear dynamics, in combining mechanistic models with empirical data, including data assimilation, and in topology. The health care process produces substantial bias, and studying that bias explicitly rather than treating it as merely another source of noise would facilitate addressing it.

Estimating summary statistics for electronic health record laboratory data for use in high-throughput phenotyping algorithms

We study the question of how to represent or summarize raw laboratory data taken from an electronic health record (EHR) using parametric model selection to reduce or cope with biases induced through clinical care. It has been previously demonstrated that the health care process (Hripcsak and Albers, 2012, 2013), as defined by measurement context (Hripcsak and Albers, 2013; Albers et al., 2012) and measurement patterns (Albers and Hripcsak, 2010, 2012), can influence how EHR data are distributed statistically (Kohane and Weber, 2013; Pivovarov et al., 2014). We construct an algorithm, PopKLD, which is based on information criterion model selection (Burnham and Anderson, 2002; Claeskens and Hjort, 2008), is intended to reduce and cope with health care process biases and to produce an intuitively understandable continuous summary. The PopKLD algorithm can be automated and is designed to be applicable in high-throughput settings; for example, the output of the PopKLD algorithm can be used as input for phenotyping algorithms. Moreover, we develop the PopKLD-CAT algorithm that transforms the continuous PopKLD summary into a categorical summary useful for applications that require categorical data such as topic modeling. We evaluate our methodology in two ways. First, we apply the method to laboratory data collected in two different health care contexts, primary versus intensive care. We show that the PopKLD preserves known physiologic features in the data that are lost when summarizing the data using more common laboratory data summaries such as mean and standard deviation. Second, for three disease-laboratory measurement pairs, we perform a phenotyping task: we use the PopKLD and PopKLD-CAT algorithms to define high and low values of the laboratory variable that are used for defining a disease state. We then compare the relationship between the PopKLD-CAT summary disease predictions and the same predictions using empirically estimated mean and standard deviation to a gold standard generated by clinical review of patient records. We find that the PopKLD laboratory data summary is substantially better at predicting disease state. The PopKLD or PopKLD-CAT algorithms are not meant to be used as phenotyping algorithms, but we use the phenotyping task to show what information can be gained when using a more informative laboratory data summary. In the process of evaluation our method we show that the different clinical contexts and laboratory measurements necessitate different statistical summaries. Similarly, leveraging the principle of maximum entropy we argue that while some laboratory data only have sufficient information to estimate a mean and standard deviation, other laboratory data captured in an EHR contain substantially more information than can be captured in higher-parameter models.

Prognostic burden of heart failure recorded in primary care, acute hospital admissions, or both: a population-based linked electronic health record cohort study in 2.1 million people

AIMS

The prognosis of patients hospitalized for worsening heart failure (HF) is well described, but not that of patients managed solely in non-acute settings such as primary care or secondary outpatient care. We assessed the distribution of HF across levels of healthcare, and assessed the prognostic differences for patients with HF either recorded in primary care (including secondary outpatient care) (PC), hospital admissions alone, or known in both contexts.

METHODS AND RESULTS

This study was part of the CALIBER programme, which comprises linked data from primary care, hospital admissions, and death certificates for 2.1 million inhabitants of England. We identified 89 554 patients with newly recorded HF, of whom 23 547 (26%) were recorded in PC but never hospitalized, 30 629 (34%) in hospital admissions but not known in PC, 23 681 (27%) in both, and 11 697 (13%) in death certificates only. The highest prescription rates of ACE inhibitors, beta-blockers, and mineralocorticoid receptor antagonists was found in patients known in both contexts. The respective 5-year survival in the first three groups was 43.9% [95% confidence interval (CI) 43.2-44.6%], 21.7% (95% CI 21.1-22.2%), and 39.8% (95% CI 39.2-40.5%), compared with 88.1% (95% CI 87.9-88.3%) in the age- and sex-matched general population.

CONCLUSION

In the general population, one in four patients with HF will not be hospitalized for worsening HF within a median follow-up of 1.7 years, yet they still have a poor 5-year prognosis. Patients admitted to hospital with worsening HF but not known with HF in primary care have the worst prognosis and management. Mitigating the prognostic burden of HF requires greater consistency across primary and secondary care in the identification, profiling, and treatment of patients.

TRIAL REGISTRATION

NCT02551016.

Comparing lagged linear correlation, lagged regression, Granger causality, and vector autoregression for uncovering associations in EHR data

Time series analysis methods have been shown to reveal clinical and biological associations in data collected in the electronic health record. We wish to develop reliable high-throughput methods for identifying adverse drug effects that are easy to implement and produce readily interpretable results. To move toward this goal, we used univariate and multivariate lagged regression models to investigate associations between twenty pairs of drug orders and laboratory measurements. Multivariate lagged regression models exhibited higher sensitivity and specificity than univariate lagged regression in the 20 examples, and incorporating autoregressive terms for labs and drugs produced more robust signals in cases of known associations among the 20 example pairings. Moreover, including inpatient admission terms in the model attenuated the signals for some cases of unlikely associations, demonstrating how multivariate lagged regression models' explicit handling of context-based variables can provide a simple way to probe for health-care processes that confound analyses of EHR data.

Predictability Bounds of Electronic Health Records

The ability to intervene in disease progression given a person's disease history has the potential to solve one of society's most pressing issues: advancing health care delivery and reducing its cost. Controlling disease progression is inherently associated with the ability to predict possible future diseases given a patient's medical history. We invoke an information-theoretic methodology to quantify the level of predictability inherent in disease histories of a large electronic health records dataset with over half a million patients. In our analysis, we progress from zeroth order through temporal informed statistics, both from an individual patient's standpoint and also considering the collective effects. Our findings confirm our intuition that knowledge of common disease progressions results in higher predictability bounds than treating disease histories independently. We complement this result by showing the point at which the temporal dependence structure vanishes with increasing orders of the time-correlated statistic. Surprisingly, we also show that shuffling individual disease histories only marginally degrades the predictability bounds. This apparent contradiction with respect to the importance of time-ordered information is indicative of the complexities involved in capturing the health-care process and the difficulties associated with utilising this information in universal prediction algorithms.

Parameterizing time in electronic health record studies

BACKGROUND

Fields like nonlinear physics offer methods for analyzing time series, but many methods require that the time series be stationary-no change in properties over time.Objective Medicine is far from stationary, but the challenge may be able to be ameliorated by reparameterizing time because clinicians tend to measure patients more frequently when they are ill and are more likely to vary.

METHODS

We compared time parameterizations, measuring variability of rate of change and magnitude of change, and looking for homogeneity of bins of temporal separation between pairs of time points. We studied four common laboratory tests drawn from 25 years of electronic health records on 4 million patients.

RESULTS

We found that sequence time-that is, simply counting the number of measurements from some start-produced more stationary time series, better explained the variation in values, and had more homogeneous bins than either traditional clock time or a recently proposed intermediate parameterization. Sequence time produced more accurate predictions in a single Gaussian process model experiment.

CONCLUSIONS

Of the three parameterizations, sequence time appeared to produce the most stationary series, possibly because clinicians adjust their sampling to the acuity of the patient. Parameterizing by sequence time may be applicable to association and clustering experiments on electronic health record data. A limitation of this study is that laboratory data were derived from only one institution. Sequence time appears to be an important potential parameterization.

Dynamical phenotyping: using temporal analysis of clinically collected physiologic data to stratify populations

Using glucose time series data from a well measured population drawn from an electronic health record (EHR) repository, the variation in predictability of glucose values quantified by the time-delayed mutual information (TDMI) was explained using a mechanistic endocrine model and manual and automated review of written patient records. The results suggest that predictability of glucose varies with health state where the relationship (e.g., linear or inverse) depends on the source of the acuity. It was found that on a fine scale in parameter variation, the less insulin required to process glucose, a condition that correlates with good health, the more predictable glucose values were. Nevertheless, the most powerful effect on predictability in the EHR subpopulation was the presence or absence of variation in health state, specifically, in- and out-of-control glucose versus in-control glucose. Both of these results are clinically and scientifically relevant because the magnitude of glucose is the most commonly used indicator of health as opposed to glucose dynamics, thus providing for a connection between a mechanistic endocrine model and direct insight to human health via clinically collected data.

Identifying and mitigating biases in EHR laboratory tests

Electronic health record (EHR) data show promise for deriving new ways of modeling human disease states. Although EHR researchers often use numerical values of laboratory tests as features in disease models, a great deal of information is contained in the context within which a laboratory test is taken. For example, the same numerical value of a creatinine test has different interpretation for a chronic kidney disease patient and a patient with acute kidney injury. We study whether EHR research studies are subject to biased results and interpretations if laboratory measurements taken in different contexts are not explicitly separated. We show that the context of a laboratory test measurement can often be captured by the way the test is measured through time. We perform three tasks to study the properties of these temporal measurement patterns. In the first task, we confirm that laboratory test measurement patterns provide additional information to the stand-alone numerical value. The second task identifies three measurement pattern motifs across a set of 70 laboratory tests performed for over 14,000 patients. Of these, one motif exhibits properties that can lead to biased research results. In the third task, we demonstrate the potential for biased results on a specific example. We conduct an association study of lipase test values to acute pancreatitis. We observe a diluted signal when using only a lipase value threshold, whereas the full association is recovered when properly accounting for lipase measurements in different contexts (leveraging the lipase measurement patterns to separate the contexts). Aggregating EHR data without separating distinct laboratory test measurement patterns can intermix patients with different diseases, leading to the confounding of signals in large-scale EHR analyses. This paper presents a methodology for leveraging measurement frequency to identify and reduce laboratory test biases.

Defining a comprehensive verotype using electronic health records for personalized medicine

The burgeoning adoption of electronic health records (EHR) introduces a golden opportunity for studying individual manifestations of myriad diseases, which is called 'EHR phenotyping'. In this paper, we break down this concept by: relating it to phenotype definitions from Johannsen; comparing it to cohort identification and disease subtyping; introducing a new concept called 'verotype' (Latin: vere = true, actually) to represent the 'true' population of similar patients for treatment purposes through the integration of genotype, phenotype, and disease subtype (eg, specific glucose value pattern in patients with diabetes) information; analyzing the value of the 'verotype' concept for personalized medicine; and outlining the potential for using network-based approaches to reverse engineer clinical disease subtypes.

Data assimilation of glucose dynamics for use in the intensive care unit

We know much about the glucose regulatory system, yet the application of this knowledge is limited because simultaneous measurements of insulin and glucose are difficult to obtain. We present a data assimilation framework for combining sparse measurements of the glucose regulatory system, available in the intensive care unit setting, with a nonlinear computational model to estimate unmeasured variables and unknown parameters. We also demonstrate a method for choosing the best variables for measurement. We anticipate that this framework will improve glucose maintenance therapies and shed light on the underlying biophysical process.