ANNs Provide Tools for Increased Diagnostic Accuracy

Medical Device & Diagnostic Industry MagazineMDDI Article IndexOriginally Published January 2000R&D HORIZONSUse of artificial neural networks is yielding medical devices that can "learn" to support care providers.

Gregg Nighswonger

January 1, 2000

14 Min Read
MDDI logo in a gray background | MDDI

Medical Device & Diagnostic Industry Magazine
MDDI Article Index

Originally Published January 2000

Use of artificial neural networks is yielding medical devices that can "learn" to support care providers.

Efforts to control rising costs, shortages of labor in a number of critical healthcare functions, and the increasing complexity of diagnostic and therapeutic systems are among the driving forces behind the ongoing development of artificial neural networks (ANNs). Computer-based diagnostic methods are being used increasingly in attempts to improve the overall quality of healthcare. Medical systems that use artificial intelligence, including expert systems and hybrid systems such as ANNs, are influencing the development of a broad range of devices. Systems of this sort are expected to dramatically improve the quality of medical diagnosis and therefore medical care.

Artificial intelligence systems that had their beginnings in the early 1970s have undergone a renaissance that began in the 1980s as computationally more powerful ANNs were developed. The latter systems used Bayesian and similar belief networks that were capable of capturing the more subtle characteristics generally displayed by clinical reasoning.1

TECHNOLOGY INSPIRED BY THE HUMAN NERVOUS SYSTEM

According to Frank Vertosick Jr., "The term neural network derives from the obvious nervous system analogy, with the processing elements serving as neurons and the connection weights equivalent to variable synaptic strengths."2 Inspired by the construction of the human nervous system, artificial neural networks are composed of interconnected devices or units, referred to as neurons, nodes, and neurodes. Like the structure and function of the nervous system and brain, ANNs generally consist of a large number of interconnected units that can process information in a highly parallel manner. Their design makes them capable not only of viewing and classifying data patterns, but of learning the classification process. With successful application to a number of diverse fields, the attraction of ANNs lies in the relative simplicity with which they can be applied to specific problems and used for performing nonlinear data processing.

"By their nature, neural networks are capable of high-speed parallel signal processing in real time," says Evangelia Micheli-Tzanakou.3 "They have an advantage over conventional technologies because they can solve problems that are too complex—problems that do not have an algorithmic solution or for which an algorithmic solution is too complex to be found." Much of the interest in developing ANN technology is because the networks can be "trained" to identify nonlinear patterns between input and output values and can solve complex problems much faster than digital computers. Micheli-Tzanakou adds that, "When used in medical diagnosis, they are not affected by factors such as human fatigue, emotional states, and habituation. They are capable of rapid identification, analysis of conditions, and diagnosis in real time."

CLASSES OF NEURAL NETWORKS

There are many different types of ANNs, with classification generally being based on how data is processed through the network. Classification of ANNs can also be based on the networks' learning method or training. Some use a supervised training method, which has been described as being similar to the student who is guided by an instructor. Other systems are classified as unsupervised or self-organizing, which has been described as being analogous to the student who learns a lesson totally on his or her own. Unsupervised ANNs cluster data into similar groups that are based on measured attributes that serve as inputs to their algorithms.

Classification is often based on the network's processing-element characteristics, the network topology, and the training or learning rules followed by the system to adapt to weighted inputs. The basic function of a neural network is to map input data in terms of its own internal connectivity. Network topology can be either feedforward (nonrecursive) or feedback (recursive). Feedforward networks have been found to offer simplicity of implementation and analysis. Although feedback systems can be applied to expand a network's range of behavior, such networks require longer training times before they can recognize inputs.

The most widely used architecture is the multilayer perceptron trained by an algorithm called backpropagation. Backpropogation is a gradient-descent algorithm that tries to minimize the average squared error of the network. In a general sense, feedforward backpropogation (FFBP) networks are input/output algorithms constructed to solve specific problems.

Such systems differ from conventional expert systems in that neural networks function without the need for human experts or ad hoc rules. Instead, training of FFBP networks entails use of a set of known input/output data points. The network alters the connection weights among its component processing elements during training to provide some stable, quantitative mapping of inputs to outputs in the training set. The trained network can then be deployed to find solutions to novel problems. The functioning of expert systems, on the other hand, is generally based on a set of a priori rules that relate the various input parameters to the output.

Vertosick states that "each approach has unique advantages and disadvantages. For example, expert systems require experts in the problem to be solved in addition to significant programming expertise. Furthermore, setting up quantitative rules is not straightforward." Describing the unique advantage of neural networks over expert systems, he adds that "neural networks . . . can be applied to problems for which no experts may yet exist."

Early uses of expert systems that provided medical reasoning included the creation of knowledge-based tools such as Casnet, Mycin, and Dialog/Internist. Each system took a distinct approach to providing the basis for interpretive decision making in a clinical setting. With the increased use of embedded computer technology, reliance on computer networks within the healthcare setting, and general advances in processing capabilities that are currently being experienced, the use of ANNs in medical devices and systems will without doubt become increasingly widespread.

INCREASING THE DIAGNOSTIC POWER OF MEDICAL DEVICES AND SYSTEMS

Increasing the accuracy of patient diagnosis remains a focal point of much of the current research involving ANN development. For example, researchers at the University of Nijmegan (Netherlands) have been examining a system that they believe will be capable of learning the interrelation of variables that characterize the state of a patient and yield the most probable diagnosis. The system also suggests optimal subsequent tests, based on the most probable diagnoses. The core of the system is a probabilistic network that is trained on the basis of patient data. Figure 1 illustrates a probabilistic network.

0001d50a.gifFigure 1. Illustration of a probabilistic network. The top layer encodes diagnoses while the bottom layer encodes findings.

The researchers note that among the principal challenges they have met was the limited number of about 400 patient records available to train the system. To overcome the problem, the group developed a method for combining the patient data set with expert knowledge provided by a physician. Use of the method allowed the system to classify 85% of test data correctly. Systems based on patient data or expert knowledge alone correctly classified about 70% of the test data. The results of the tests suggested that the network's diagnostic and selection performance is comparable with the performance of experienced physicians.

The complex environment of a hospital intensive-care unit poses critical challenges to the application of ANN technology. Large amounts of data must be entered into charts by nursing staff in addition to monitoring patient parameters and administering treatment. Physicians must then interpret the significant volume of data represented by these activities in order to render clinical judgment about each patient's condition and plan ongoing treatment.

Researchers at the University of Aberdeen (Scotland) have collaborated with Kontron Instruments (Watford, Herts, UK) to develop advanced information-handling systems for use in intensive-care units. The UNIX-based system incorporates automatic data acquisition and storage, graphical user interfaces at work-stations, and intelligent decision support. The integrated decision support tools are expected to be used on a regular basis in routine clinical data processing, according to the researchers. A critiquing expert system has also been developed that is capable of commenting on prescriptions, and the group is working on systems for continuous analysis of cardiac data as the basis for alarm and state interpretation. Long range goals of the research include integration of separate knowledge-based systems to increase capabilities for summarizing the overall patient state.

IMPROVING THE ACCURACY OF ROUTINE PAP SCREENING

Researchers have noted that successful clinical decision systems are designed to support human professionals in routine clinical tasks. In essence, the computer does not replace the entire decision-making process—only parts of it. Use of such systems in Pap test screening provides a good example.

A number of automated systems based on ANN technology have been approved by FDA to improve the accuracy of Pap tests. The Pap test has, of course, become a routine part of gynecological exams, and was largely responsible for the 70% reduction between 1950 and 1970 in the number of women dying from cervical cancer, according to the National Cancer Institute. Pap screening methods, however, have been prone to error. In addition to the labor-intensive nature of the screening process, which can increase the likelihood of human error, cell abnormalities can be difficult to detect even by the trained eye or can be masked by infection.

Advanced systems for both slide preparation and screening have allowed Pap testing to become largely automated and have resulted in marked increases in accuracy. The PAPNET system, which incorporates neural net technology, was based on technology developed originally as part of the "Star Wars" missile defense research program. The system uses an ANN-based computer system that has been trained by example to detect abnormal cervical cells. The system essentially learned how to screen slides by being fed a series of digitized images of Pap slides, allowing the system to generate guidelines to abnormal cell detection.

PAPNET screens each slide that cytotechnologists have identified as "normal" and selects the 128 cells or cell clusters that are most likely to be abnormal. The cytotechnologist then rescreens enlarged color images of these cells. Reexamining negative Pap slides that were taken from women with high-grade cervical cell abnormality or cervical cancer has shown that PAPNET testing detected about one-third of the abnormalities missed by manual screening on previous Pap smears.

The computerized AutoPap 300 QC Pap test rescreening system uses high-speed video microscopy and proprietary algorithms to interpret images and classify slides. One study comparing the system with conventional rescreening techniques found that cytotechnologists who randomly rescreened 10% of more than 4000 slides originally classified as "normal" detected only 1 of every 10 false negatives present. Using the AutoPap system to select a 10% sampling of slides that it classified as most abnormal, the cytotechnologists detected up to half of all missed abnormal samples.

Because early detection of precancerous cells remains the most effective means of preventing cervical cancer, continued development of automated screening and rescreening systems is expected to enhance the diagnostic value of the Pap test.

CORONARY ARTERY DISEASE

Greater reliance by hospitals on increasingly powerful computer information systems and data storage in digital formats, combined with the use of ANN-based systems, is providing the foundation for advanced methods of diagnosing and treating coronary artery disease (CAD). Researchers at Pacific Northwest National Laboratory (PNNL) have conducted preliminary research to assess the effectiveness of ANNs designed to aid in detecting CAD.

Use of single-photon emission computed tomography (SPECT) has been proven to be a particularly useful imaging modality for CAD detection. SPECT operates by collecting a series of two-dimensional scintigraphic images of a patient's body. The initial images are used to generate a three-dimensional model of the chest. An algorithm is then applied to the model to create a two-dimensional polar plot of those cardiac regions that are of particular interest to the physician in diagnosing CAD.

The PNNL researchers are exploring the use of ANN technology that takes advantage of SPECT's digital format and the processing power of current computer systems to enhance conventional diagnostic methods. The two-dimensional images of the chest were used to develop an algorithm capable of extracting the region surrounding the patient's heart in each image. This subsequent set of images was then reduced and prepared as input for the ANN.

The researchers trained the ANN using data on 31 patients, including 16 healthy subjects and 15 with CAD. The group then used a disjoint set of patients, including five healthy and five diseased subjects, as a validation set to determine when training was to be stopped. The network was found to be capable of correctly diagnosing 14 of the 15 CAD subjects and 14 of the 16 healthy subjects; it also correctly classified all patients in the validation set. When a second disjoint group of patients was used, the network correctly classified four of seven patients with CAD and four of six healthy patients. The PNNL group found that almost all of the CAD patients in the training and validation sets suffered from disease in the heart's inferolateral wall. The four correctly classified patients in the final test group also suffered from disease in the inferolateral wall. The other patients had disease in other regions of the heart.

The researchers determined that the network had been trained only to identify disease in the one region of the heart; disease in other regions confused the ANN. The lack of presented patients with disease in other coronary regions meant that the ANN could not learn by example.

The PNNL group concluded that it "is indeed feasible to train an ANN that can diagnose CAD." The researchers have been collecting additional data to retrain the ANN. A series of related projects, labeled radiology diagnostics using artificial neural networks (RADANN), has been initiated by the PNNL group. The series involves the diagnosis of various disease conditions using radiological images to train ANN-based diagnostic tools. PNNL is seeking partners to support further development of these systems.

INTELLIGENT ALARMS FOR MONITORS

Use of ANNs has also been applied to alarm components of various patient-monitoring systems. Adding intelligence to monitor alarms is serving to make them more dependable and effective. Hospital studies have found, for example, that monitor alarm sensitivity can have a surprising impact on patient safety when human factors are considered. Alarm thresholds that are capable of triggering in instances where they are truly needed will also trigger a certain number of false alarms. Alarms with thresholds that allow an excessive number of false alarms, however, have been found to promote human error if they prompt hospital staff to turn off the alarms rather than reset them.

One example of monitoring systems that can be enhanced through the use of ANNs is cardiotocograms (CTGs) used routinely for fetal monitoring before birth. One problem with CTG methods has been the need to check frequently for paper output produced by the device. Researchers at the University of Vienna (Austria) have found that when CTG traces are checked automatically, the interval between a dangerous situation and an intervention can be shortened. Most attempts to automate monitoring tools for paper output have failed, however, because they are not reliable enough.

The researchers have examined the application of neural networks to creating a more reliable method of monitoring the CTG. They note that they can be applied to a number of subtasks within the monitoring process. The researchers have focused primarily on the detection of deceleration because this process provides a major source of information. They also examined the application of neural networks to forecasting fetal heartbeat intervals, which has been found to be useful for identifying CTG artifacts.

The approach taken by the Vienna researchers focused on recording single heartbeat intervals, which provided the basis for calculating all parameters. The group compared sample CTG traces tested with a number of conventional methods that used averaged values, and the neural-network-based system. The various methods were used to detect decelerations in the CTG streams. To ensure adequate sampling, numerous data sets were collected, interpreted, and marked by a specialist. The systems were then used to classify the samples.

The outputs of the trained neural networks were found to indicate that a deceleration is "strongly dependent on the variability of the fetal heart rate," suggesting that exact measures of long- and short-term variability are critical. They suggested that various measures of variability could be integrated in future systems to increase correct classifications.

CONCLUSION

Healthcare has provided a profoundly stimulating environment for the development of artificial intelligence systems, and particularly ANN-based programs. Continuing research is providing the foundation for systems that can learn more quickly and effectively, supporting clinical diagnoses that are more accurate. And new computer technologies are making it possible to make such systems an integral component in advanced medical devices. From assessment of coronary artery disease and Pap smear screening to patient assessment in the ICU and other clinical settings, ANNs are increasingly being put to work to support the efforts of clinicians and enhance the level of care.

REFERENCES

1. Casimir A Kulikowski, "History and Development of Artificial Intelligence Methods for Medical Decision Making," in The Biomedical Engineering Handbook, ed. Joseph D Bronzino (Boca Raton, FL: CRC Press, 1995), 2681.

2. Frank Vertosick Jr., "Neural Networks," in Handbook of Clinical Automation, Robotics, and Optimization, ed. Gerald J Kost with collaboration of Judith Welsh (New York: Wiley, 1996) 79—91.

3. Evangelia Micheli-Tzanakou, "Neural Networks in Biomedical Signal Processing," in The Biomedical Engineering Handbook, ed. Joseph D Bronzino (Boca Raton, FL: CRC Press, 1995), 917.

Gregg Nighswonger is executive editor of MDDI.


Return to the MDDI January table of contents | Return to the MDDI home page

Copyright ©2000 Medical Device & Diagnostic Industry

Sign up for the QMED & MD+DI Daily newsletter.

You May Also Like