NEURAL-NETWORK PREDICTION OF NONSTRESS TEST-RESULTS - HOW OFTEN SHOULD WE PERFORM NONSTRESS TESTS

OBJECTIVE: Our purpose was to predict outcomes and optimal intervals f or nonstress tests of term gravid women with neural networks. STUDY DE SIGN: We studied 100 normal term patients whose 30-minute nonstress te sts, performed on 5 consecutive days, were computer analyzed for the f ollowing elements: fetal heart rate baseline, variability, signal loss , accelerations (> 15 beats/min), and decelerations. The training set used 65 patients; the testing, 35 patients. Nonstress test data (days 1 to 4) were inputs; day 5 data were training patterns. Networks for e ach nonstress test element used Brainmaker Macintosh 1.0 (California S cientific Software, Nevada City, Calif.) trained to 0.12 tolerance. Ac tual fetal heart rate elements and their daily differences were compar ed with predictions by the networks and multiple regressions. RESULTS: There was little difference between networks using daily or. alternat e-day inputs for predicting test performance on day 5; networks using test intervals > 2 days could not be trained to tolerance. Long-term f etal heart rate variation was the nonstress test element best predicte d. Daily differences networks provided better prediction of all day 5 data than did actual daily values networks or multiple regression form ulas. CONCLUSIONS: Baseline long-term fetal heart rate variability see ms to be the most predictable fetal heart rate element over time and s hould merit more consideration in overall fetal testing. Fetal heart r ate elements are not easily predicted by any method for intervals long er than 2 days. Using longer test intervals might run a greater risk f or unanticipated changes in nonstress test outcomes, even when fetal c ondition is normal.