DPC Symposium Attendees "Plug In" to
Neural Networks

In the clinical laboratory, computer expert systems can streamline test ordering and improve efficiency.¹ Expert systems incorporate simple decision-trees for testing schemes, as well as more complex statistical analyses. DPC's Luncheon Symposium, held during the AACC's 1999 National Meeting, featured neural networks as one such analytical tool. The speakers, Mr. Jim DeLeo and Dr. Alan Remaley, work with neural networks in the clinical laboratory at the National Institutes of Health. They discussed successful medical applications of neural networks, focusing particularly on monitoring the quality control of automated analyzers using patient-based laboratory data.

Mr. DeLeo, a software developer who has worked in the biomedical sector for over 20 years, began the symposium by describing biomedical computing methods. "Smart computing" methods include machine learning, biology-based computing, fuzzy logic, artificial intelligence, statistical methodology, cluster analysis, and ROC methodology. Various programs have been developed to facilitate the interpretation of laboratory results, but the question arises whether the methods improve the ability of clinical tests to discriminate among medical conditions. Receiver operating characteristic (ROC) plots answer this question. Originally used in World War II to evaluate radar performance in discriminating warplanes and missiles from other flying objects, ROC methodology is now widely used in clinical medicine to evaluate assays by graphically comparing assay sensitivities and specificities.

Mr. DeLeo described neural networks as computing paradigms inspired by the neurosciences. The algorithms use interconnected networks of simple computing elements, modeled after a biological nervous system comprised of highly interconnected neurons. The programs read input data from various sources—clinical parameters and laboratory data, for instance— connect the input information via computing nodes, and generate an output signal. The iterative computing process refines the output signal to best match the outcome given in the training data set, highlighting the importance of working with good data. The weighted relationships among the inputs depend solely on the characteristics of the training data, not on preconceived notions about the input relationships. Neural networks have great objective power, but they depend absolutely on reliable data to generate a robust algorithm.

Mr. DeLeo also described neural network applications in the assessment of cardiac risk. He showed the potential patient-care benefits of cardiac risk vs. time plots and the effect of behavioral changes on the risk assessment. Another current project performs online instrument quality control. Analytes that show bias variance errors on the machine are flagged to the operator.

In contrast to a computer scientist's point of view, Dr. Remaley gave a clinician's perspective. He described computer applications in medicine, including data management, communication, instrument operation, and clinical decision support (CDS). CDS systems show promise for improving productivity, quality, and standardization. As biomedical literature expands exponentially, keeping abreast of developments becomes increasingly difficult and requires data management tools. Medical decision-making entails both analytical and empirical thought processes, which neural networks approximate.

Neural networks have been successfully applied to many areas in medicine, including diagnosis, image processing, waveform analysis, outcome prediction, and histology. Dr. Remaley described a study by Baxt et al.,² which demonstrated the effectiveness of neural networks in diagnosing myocardial infarction. The network had a sensitivity and specificity of 97% and 96%, respectively, compared to 78% and 85% for physicians diagnosing the same patients in the ER. Cardiac risk-factor analysis may also benefit from computational programs. By combining inputs as general as age and gender with more specific variables, such as blood pressure and cholesterol levels, neural networks may be able to diagnose patients at high risk.

Dr. Remaley concluded with a discussion of some obstacles to implementing neural networks in the laboratory, which include hardware, software and personnel barriers. The "black box" barrier refers to people's reluctance to adopt unfamiliar technology. Dr. Remaley and Mr. DeLeo described computer aids in medicine and listed specific successes. We hope this seminar helps to overcome the "black box" barrier and promote the implementation of computational tools in the clinical laboratory.

References

1. Smith BJ, McNeely MD. The influence of an expert system for test ordering and interpretation on laboratory investigations. Clin Chem 1999;45:1168-75.

2. Baxt WG, Skora J. Prospective validation of artificial neural network trained to identify acute myocardial infarction. Lancet 1996;347:12-5.