Merging medical informatics and automated diagnostic methods.

In many instances disease diagnosis is more of an art than a science due to the complexity of disease, lack of detailed information on parameters that are indicative of the disease, and lack of sufficient data to apply these parameters to both diagnosis and treatment. Broad-based expansion of electronic health records (EHRs) will produce additional data for improved model development. However many obstacles remain. Patient record content is not broadly available because of privacy concerns and the lack of standardization of EHR formats. If available on a large scale, de-identified medical records can provide a basis for development of disease models by removing privacy concerns. Once comprehensive disease models have been developed that assist in identifying possible diseases and also include parameters that were utilized along with their relative importance, automated analytic methods can be used to indicate the likelihood of the presence of specific diseases. Although the physician will always remain as the final expert, these methods can provide an expanded information set and provide analysis that is too complex for standard methods.

Overcoming barriers to development of cooperative medical decision support models.

Attempts to automate the medical decision making process have been underway for the at least fifty years, beginning with data-based approaches that relied chiefly on statistically-based methods. Approaches expanded to include knowledge-based systems, both linear and non-linear neural networks, agent-based systems, and hybrid methods. While some of these models produced excellent results none have been used extensively in medical practice. In order to move these methods forward into practical use, a number of obstacles must be overcome, including validation of existing systems on large data sets, development of methods for including new knowledge as it becomes available, construction of a broad range of decision models, and development of non-intrusive methods that allow the physician to use these decision aids in conjunction with, not instead of, his or her own medical knowledge. None of these four requirements will come easily. A cooperative effort among researchers, including practicing MDs, is vital, particularly as more information on diseases and their contributing factors continues to expand resulting in more parameters than the human decision maker can process effectively. In this article some of the basic structures that are necessary to facilitate the use of an automated decision support system are discussed, along with potential methods for overcoming existing barriers.

Uncertainty and complexity in personal health records.

New technologies in medicine have led to an explosion in the number of parameters that must be considered when diagnosing and treating a patient. Because of this high volume of data it is not possible for the human decision maker to take all information into account in arriving at a decision. Automated methods are needed to effectively evaluate electronic information in many formats and provide summaries to the medical professional. The task is complicated by the complexity of the data and the potential uncertainty of some of the results. In this article complexity and uncertainty in medical data are discussed in terms of both representation and types of analysis. Methods that can address multiple complex data types are illustrated and examples are provided for specific medical problems. These methods are particularly important for automated trend analysis in the personal health record as small errors can be propagated through the complex system resulting in incorrect diagnosis and treatment.

Diagnostic models based on personalized analysis of trends (PAT).

Many changes have taken place in medicine over the last century. In the first-half of the 20th century physicians were faced with the challenge of making diagnoses with too little information, often resorting to exploratory surgery to confirm the presence or absence of a condition. Due to rapid technological advances during the second-half of the 20th century, and continuing to this day, the position of the physician has now shifted from an information-poor environment to an environment with too much information, often exceeding the limits of human decision-making capabilities. To take full advantage of all available information, a new approach based on refined automated decision support methods is needed to assist the physician in the decision-making process. Medical decision support systems need to evolve from stand-alone systems to cooperative systems in which the physician becomes the decision maker, but relies on the decision support system to sift through information to determine relevant trends. In this paper, a decision support system that combines a number of methodologies for trend analysis is described, along with examples in cardiology. The methods have also been used in applications in neurology as well as cancer diagnosis and prognosis.

Multidimensional medical decision making.

A century ago the physician had too little information on which to determine accurate diagnoses. Due to the rapid progression of technology in the Twentieth Century and the beginning of the Twenty-First Century, this situation changed significantly. Now the physician is faced with multi-parameter analyses that include sophisticated imaging, advanced cardiovascular studies, extensive laboratory tests, and genetic information, all of which impact diagnosis, treatment, and prognosis. New informatics tools are needed to assist, not replace, the physician in the decision process. Decision analysis tools must be flexible to accommodate new methods of diagnosis as well as advances in information technology. In this article, basic structures are defined that can form the basis of such a system.

Temporal trend analysis in personal health records.

The creation of a universal standard for electronic medical records (EMR) remains a work in progress. The existence of a standard EMR that can be accessed by medical professionals anywhere in the world as well as by the patient will have a significant impact on the way that medicine is practiced. In addition to having information readily available, it would present the possibility to do individual trend analysis on each patient based on information accumulated over a lifetime. This information would permit diagnosis based on the individual rather than solely on population statistics. A comprehensive system for performing trend analysis on the multitude of data types relevant to medical decision making requires the creation and integration of a number of methods. A general approach based on a hybrid combination of automated decision support methods is described and illustrated for using trend data applied to cardiac diagnosis. The methods are general and can be applied to any application area.

Combining evidence in hybrid medical decision support models.

Hybrid methods are particularly useful for building diagnostic models based on biomedical data due to the wide variety of data types that are routinely encountered. Evaluation of the effectiveness of hybrid systems is complicated when multiple methods are combined to reach a conclusion. In the work described here, methods for combining results based on the general reliability of each model as well as its applicability to the case under evaluation are presented. Reliability measures differ depending on whether symbolic or numeric information is analyzed and depend on the strength of the decision algorithm as well as the soundness of the domain knowledge upon which the decision is based. In addition to reliability, combination of results is complicated by the need to weight each method to form the final conclusion. Weighting factors depend on the degree of certainty that the decision is correct for each of the methods. The process is illustrated in an application to cardiac diagnosis.

Synchronization measures of the scalp electroencephalogram can discriminate healthy from Alzheimer's subjects.

Three synchronization measures are applied to scalp electroencephalogram (EEG) data collected from 20 patients diagnosed to have either: (1) no dementia, (2) mild cognitive impairment (MCI), or (3) Alzheimer's disease (AD). We apply the three synchronization measures--the phase synchronization, and two measures of nonlinear interdependency--to the data collected from awake patients resting with eyes closed. We show that the synchronization in potential between electrodes near the left and right occipital lobes provides a statistically significant discriminant between the healthy and AD subjects, and the MCI and AD subjects. None of the three measures appears able to distinguish between the healthy and MCI subjects, although MCI subjects show synchronization values intermediate between healthy subjects (with high synchronization values) and AD subjects (with low synchronization values) on average.

Neural network models for biosignal analysis.

Signal analysis provides important clues for diagnosis of disease in many arenas, particularly in cardiology. While each result may give good diagnostic information, a comprehensive decision model requires the combination of results. In the work described here, a neural network model is used to combine various features obtained through signal analysis. The original model is based on electrocardiogram (ECG analysis). This model is extended in two ways: the neural network model is expanded to include other clinical parameters in addition to the ECG and the method is generalized to include other biosignals.

Intelligent analysis of biosignals.

While biosignal analysis has been a mainstay in many medical applications for a number of decades, notably in cardiology, problems remain in the analysis of high volume data and in comparison among time series. Effective comparisons can determine if the state of a particular patient has changed significantly. These comparisons are useful in both long-term and short-term monitoring. Methods will vary depending on the type of signal and the situation. Intelligent agents permit the activation of the appropriate analysis for the signal and the current situation. Agents are defined in terms of both methodology and function. The agent system is described and illustrated in applications to cardiology and neurology.

Diagnostic potential of nonlinear analysis of biosignals.

Biosignals have played an important role in medical diagnosis. The first biosignal to be used extensively was the electrocardiogram whose interpretation initially relied on manual analysis of paper tracings. Interpretation was based on variations of the normal QRS pattern associated with each heartbeat. Automated arrhythmia analysis was developed commercially and has been in standard clinical use for some time. The advent of Holter monitoring presented new challenges for the analysis of very long time series. New methods have been developed for this purpose, including nonlinear dynamical approaches. These methods have yielded important diagnostic clues. In this article, the diagnostic use of parameters derived from nonlinear analysis, both alone and in conjunction with other clinical information, is discussed.

Prediction of traumatic wound infection with a neural network-derived decision model.

The objective of this study was to develop and validate a decision model, using an artificial neural network, that predicts infection in uncomplicated, traumatic, sutured wounds. The study was a prospective, cohort study of all patients presenting to the emergency department of a county teaching hospital with uncomplicated wounds that required suturing. In evaluating and treating wounds, emergency medicine (EM) faculty and residents, resident physicians in primary-care specialties, and supervised medical students on EM clerkships followed a standardized wound-management protocol. Clinicians estimated the likelihood of subsequent infection using a 5-point scale. Wound healing was followed until sutures were removed. Wound outcome data were collected by medical personnel blinded to the initial prediction. Student's t-tests and Pearson's chi-square statistic were used to identify independent predictors that served as input variables. Wound infection was the single output variable. Neural network analysis was used to assign weights to input variables and derive a decision equation. A total of 1,142 wounds were analyzed in the study. The overall infection rate was 7.2%. The most predictive factors for wound infection were wound location, wound age, depth, configuration, contamination, and patient age. To derive a decision equation for the model, the network was trained on data from half of the subjects and tested on the remainder. When used as a diagnostic test for wound infection, the decision model had a sensitivity of 70%, as compared to 54% for physicians, and a specificity of 76%, as compared to 78% for physicians. We conclude that through the use of combinations of 7 clinical variables available at the time of initial wound management, a neural network-derived decision model may be used to identify uncomplicated, traumatic wounds at higher risk for infection.