Learning temporal probabilistic causal models from longitudinal data

Prediction model for glucose metabolism based on lipid metabolism

OBJECTIVES

We developed a robust, long-term clinical prediction model to predict conditions leading to early diabetes using laboratory values other than blood glucose and insulin levels. Our model protects against missing data and noise that occur during long-term analysis.

METHODS

RESULTS of a 75-g oral glucose tolerance test (OGTT) were divided into three groups: diabetes, impaired glucose tolerance (IGT), and normal (n = 114, 235, and 325, respectively). For glucose metabolic and lipid metabolic parameters, near 30-day mean values and 10-year integrated values were compared. The relation between high-density lipoprotein cholesterol (HDL-C) and variations in HbA1c was analyzed in 158 patients. We also constructed a state space model consisting of an observation model (HDL-C and HbA1c) and an internal model (disorders of lipid metabolism and glucose metabolism) and applied this model to 116 cases.

RESULTS

The root mean square error between the observed HbA1c and predicted HbA1c was 0.25.

CONCLUSIONS

In the observation model, HDL-C levels were useful for prediction of increases in HbA1c. Even with numerous missing values over time, as occurs in clinical practice, clinically valid predictions can be made using this state space model.

Using hierarchical dynamic Bayesian networks to investigate dynamics of organ failure in patients in the Intensive Care Unit

In intensive care medicine close monitoring of organ failure status is important for the prognosis of patients and for choices regarding ICU management. Major challenges in analyzing the multitude of data pertaining to the functioning of the organ systems over time are to extract meaningful clinical patterns and to provide predictions for the future course of diseases. With their explicit states and probabilistic state transitions, Markov models seem to fit this purpose well. In complex domains such as intensive care a choice is often made between a simple model that is estimated from the data, or a more complex model in which the parameters are provided by domain experts. Our primary aim is to combine these approaches and develop a set of complex Markov models based on clinical data. In this paper we describe the design choices underlying the models, which enable them to identify temporal patterns, predict outcomes, and test clinical hypotheses. Our models are characterized by the choice of the dynamic hierarchical Bayesian network structure and the use of logistic regression equations in estimating the transition probabilities. We demonstrate the induction, inference, evaluation, and use of these models in practice in a case-study of patients with severe sepsis admitted to four Dutch ICUs.

Practical issues in the identification of empirical models from simulated type 1 diabetes data

BACKGROUND

A model-based controller for an artificial beta-cell automatically regulates blood glucose levels based on available glucose measurements, insulin infusion and meal information, and model predictions of future glucose trends. Thus, the identification of simple, accurate models plays an important role in the development of an artificial beta-cell.

METHODS

Glucose data simulated from a nonlinear physiological model of type 1 diabetes are used to identify linear dynamic models of two types: autoregressive exogenous input (ARX) and output-error (OE) models. The model inputs are meal carbohydrates and exogenous insulin, which in practice are often administered simultaneously and in the same ratio, i.e., the insulin-to-carbohydrate ratio. The effect of modeling these inputs as impulses versus time-smoothed profiles ("transformed inputs") is explored in depth. The models are evaluated based on their ability to describe the data from which they were identified (i.e., calibration data) as well as independent data (i.e., validation data).

RESULTS

In general, the best models described their calibration data more accurately using transformed inputs (R(Cal) (2) = 71% for the ARX models and R (Cal) (2) = 78% for the OE models) than using impulse inputs (R (Cal) (2) = 14% for the ARX models and R (Cal) (2) = 70% for the OE models). The only model/input combination that resulted in consistently accurate validation fits was the ARX models using transformed inputs (39% <or= R (Val) (2) <or= 58%).

CONCLUSIONS

When identifying non-physiologically based models from diabetes data with simultaneous and proportional meals and insulin boluses, model accuracy is improved by modeling the inputs as time-smoothed profiles. Also, while OE models describe their calibration data very well, ARX models more accurately describe validation data. Their versatility makes ARX models a more attractive choice for implementation in a model-based controller of an artificial beta-cell.

A hierarchical Naïve Bayes Model for handling sample heterogeneity in classification problems: an application to tissue microarrays

Background

Uncertainty often affects molecular biology experiments and data for different reasons. Heterogeneity of gene or protein expression within the same tumor tissue is an example of biological uncertainty which should be taken into account when molecular markers are used in decision making. Tissue Microarray (TMA) experiments allow for large scale profiling of tissue biopsies, investigating protein patterns characterizing specific disease states. TMA studies deal with multiple sampling of the same patient, and therefore with multiple measurements of same protein target, to account for possible biological heterogeneity. The aim of this paper is to provide and validate a classification model taking into consideration the uncertainty associated with measuring replicate samples.

Results

We propose an extension of the well-known Naïve Bayes classifier, which accounts for biological heterogeneity in a probabilistic framework, relying on Bayesian hierarchical models. The model, which can be efficiently learned from the training dataset, exploits a closed-form of classification equation, thus providing no additional computational cost with respect to the standard Naïve Bayes classifier. We validated the approach on several simulated datasets comparing its performances with the Naïve Bayes classifier. Moreover, we demonstrated that explicitly dealing with heterogeneity can improve classification accuracy on a TMA prostate cancer dataset.

Conclusion

The proposed Hierarchical Naïve Bayes classifier can be conveniently applied in problems where within sample heterogeneity must be taken into account, such as TMA experiments and biological contexts where several measurements (replicates) are available for the same biological sample. The performance of the new approach is better than the standard Naïve Bayes model, in particular when the within sample heterogeneity is different in the different classes.

A computer system for interpreting blood glucose data

This paper presents an overview on the design and implementation of a computer system for the interpretation of home monitoring data of diabetic patients. The comprehensive methodology covers the major information processing steps leading from raw data to a concise summary of what has happened between two subsequent visits. It includes techniques for summarising and interpreting data, checking for inconsistency, identifying and diagnosing metabolic problems and learning from patient data. Data interpretation focuses on extracting trend patterns and classifying/clustering daily blood glucose (BG) profiles. The software helps clinicians to explore data recorded before the main meals and bedtime, and to identify problems in the patient's metabolic control which should be addressed either by educating the patient and/or adjusting the current management regimen.

Integrating model-based decision support in a multi-modal reasoning system for managing type 1 diabetic patients

We present a multi-modal reasoning (MMR) methodology that integrates case-based reasoning (CBR), rule-based reasoning (RBR) and model-based reasoning (MBR), meant to provide physicians with a reliable decision support tool in the context of type 1 diabetes mellitus management. In particular, we have implemented a decision support system that is able to jointly exploit a probabilistic model of the glucose-insulin system at the steady state, a RBR system for suggestion generation and a CBR system for patient's profiling. The integration of the CBR, RBR and MBR paradigms allows for an optimized exploitation of all the available information, and for the definition of a therapy properly tailored to the patient's needs, overcoming the single approaches limitations. The system has been tested both on simulated and on real patients' data.

A multi-modal reasoning methodology for managing IDDM patients

We present a knowledge management and decision support methodology for insulin dependent diabetes mellitus (IDDM) patients care. Such methodology exploits the integration of case based reasoning (CBR) and rule based reasoning (RBR), with the aim of helping physicians during therapy planning, by overcoming the intrinsic limitations shown by the independent application of the two reasoning paradigms. RBR provides suggestions on the basis of a situation detection mechanism that relies on formalized prior knowledge; CBR is used to specialize and dynamically adapt the rules on the basis of the patient's characteristics and of the accumulated experience. When the case library is not representative of the overall population, only RBR is applied to define a therapy for the input situation, which can then be retained, enriching the case library competence. The paper reports the first evaluation results, obtained both on simulated examples and on real patients. This work was developed within the EU funded telematic management of insulin dependent diabetes mellitus (T-IDDM) project, and is fully integrated in its web-based architecture.

Diabetic patients management exploiting case-based reasoning techniques

In this paper we propose a case-based decision support tool, designed to help physicians in 1st type diabetes therapy revision through the intelligent retrieval of data related to past situations (or 'cases') similar to the current one. A case is defined as a set of variable values (or features) collected during a visit. We defined taxonomy of prototypical patients' conditions, or classes, to which each case should belong. For each input case, the system allows the physician to find similar past cases, both from the same patient and from different ones. We have implemented a two-steps procedure; (1) it finds the classes to which the input case could belong; (2) it lists the most similar cases from these classes, through a nearest neighbor technique, and provides some statistics useful for decision taking. The performance of the system has been tested on a data-base of 147 real cases, collected at the Policlinico S. Matteo Hospital of Pavia. The tool is fully integrated in the web-based architecture of the EU funded Telematic management of Insulin Dependent Diabetes Mellitus (T-IDDM) project.

Dynamic updating in DIAS-NIDDM and DIAS causal probabilistic networks

Diabetes advisory system (DIAS) is a decision support system, which has been developed to provide advice on the amount of insulin injected by subjects with insulin-dependent diabetes mellitus (IDDM). DIAS employs a temporal causal probabilistic network (CPN) to implement a stochastic model of carbohydrate metabolism. The CPN network has recently been extended to provide also advice to subjects with noninsulin-dependent diabetes mellitus (NIDDM). However, due to increased complexity and size of the extended CPN the calculations became unfeasible. The CPN network was, therefore, simplified and a novel approach employed to generate conditional probability tables. The principles of dynamic CPN's were adopted and, in combination with the method of conditioning, learning, and forecasting, were implemented in a time- and memory-efficient way. An evaluation using experimental data was carried out to compare the original and revised DIAS implementations employing data collected by patients with IDDM, and to assess the a posteriori identifiability of model parameters in patients with NIDDM.

Modelling behavioral syndromes using Bayesian networks

In this paper Bayesian networks modelling is applied to a multidimensional model of depression. The characterization of the probabilistic model exploits expert knowledge to associate latent concentrations of neurotransmitters and symptoms. An evolution perspective is also considered. Specific criteria are introduced to detect the influence of the latent variable on the observation of symptoms. The Bayesian analysis is carried out using Gibbs sampling technique which is implemented in the BUGS software. The estimation phase leads to the selection of symptoms entering into the definition of behavioral syndromes. Results on real data are discussed. The last section deals with simulation experiments. Simulation results confirm our methodological choices. Results of the paper can enlarge to the central problem of the management of latent variables in Bayesian networks modelling.

A distributed system for diabetic patient management

This paper describes a telemedicine system for diabetic patients management, presenting its architecture, the technical solutions adopted and the methodologies on which it is based. The system, designed to provide decision support in a distributed environment, is composed of two modules, a Patient Unit and a Medical Unit, connected by telecommunication services. We outline how the two modules can interact to perform an effective monitoring and a cooperative control of glucose metabolism. In particular, we detail the data analysis tasks performed by the two units and how the results are exploited to assist patients and physicians in revising and adjusting the therapeutic protocol. We will finally describe the current prototypical implementation of the system that uses HTTP as the communication protocol and HTML pages as the graphical user interface.

Effective data validation of high-frequency data: time-point-, time-interval-, and trend-based methods

Real-time systems for monitoring and therapy planning, which receive their data from on-line monitoring equipment and computer-based patient records, require reliable data. Data validation has to utilize and combine a set of fast methods to detect, eliminate, and repair faulty data, which may lead to life-threatening conclusions. The strength of data validation results from the combination of numerical and knowledge-based methods applied to both continuously-assessed high-frequency data and discontinuously-assessed data. Dealing with high-frequency data, examining single measurements is not sufficient. It is essential to take into account the behavior of parameters over time. We present time-point-, time-interval-, and trend-based methods for validation and repair. These are complemented by time-independent methods for determining an overall reliability of measurements. The data validation benefits from the temporal data-abstraction process, which provides automatically derived qualitative values and patterns. The temporal abstraction is oriented on a context-sensitive and expectation-guided principle. Additional knowledge derived from domain experts forms an essential part for all of these methods. The methods are applied in the field of artificial ventilation of newborn infants. Examples from the real-time monitoring and therapy-planning system VIE-VENT illustrate the usefulness and effectiveness of the methods.

Temporal reasoning and temporal data maintenance in medicine: issues and challenges

We present a brief, nonexhaustive overview of research efforts in designing and developing time-oriented systems in medicine. The growing volume of research on time-oriented systems in medicine can be viewed from either an application point of view, focusing on different generic tasks (e.g. diagnosis) and clinical areas (e.g. cardiology), or from a methodological point of view, distinguishing between different theoretical approaches. In this overview, we focus on highlighting methodological and theoretical choices, and conclude with suggestions for new research directions. Two main research directions can be noted: temporal reasoning, which supports various temporal inference tasks (e.g. temporal abstraction, time-oriented decision support, forecasting, data validation), and temporal data maintenance, which deals with storage and retrieval of data that have heterogeneous temporal dimensions. Efforts common to both research areas include the modeling of time, of temporal entities, and of temporal queries. We suggest that tasks such as abstraction of time-oriented data and the handling of different temporal-granularity levels should provide common ground for collaboration between the two research directions and fruitful areas for future research.

Temporal diagnostic reasoning based on time-objects

Time is essential in diagnostic problem-solving. However, as with other commonsense tasks, time representation and reasoning is not a trivial undertaking. This probably explains why time has either been ignored or implicitly represented and used in the majority of diagnostic systems, medical or otherwise. Durations, temporal uncertainty and multiple temporal granularities are necessary requirements for medical problem-solving. Most general theories of time proposed in the literature do not address all these requirements, and some do not address any. The paper discusses time representation and reasoning in medical diagnostic problem-solving, building from a generic temporal ontology which covers the above temporal requirements. Much of what is discussed, however, is applicable to non-medical domains as well. It is argued that the diagnostic concepts (patient data, disorders, therapeutic-actions) are naturally modelled as time-objects. The resulting representation treats time as an integral dimension to these concepts, with special status. Time-object-based representations for generic hypotheses (disorders, actions) are discussed and illustrated; in the case of disorders the representation covers both an associational model and a causal-associational model. A central function of diagnostic problem-solving is deciding the compatibility of hypotheses with regard to a patient model. In this respect the paper discusses temporal and contextual screening of triggered hypotheses as well as accountings and conflicts between time-objects.