Semicontinuous cardiac output monitoring using a neural network.

OBJECTIVES: This study compared 2-mL bolus thermodilution cardiac output measurements with standard 10-mL bolus measurements. DESIGN: Cardiac output was measured with the new 2-mL bolus technique and the 10-mL standard thermodilution technique in a perspective series. We describe a system that automatically cools and injects 2-mL boluses of saline into a standard pulmonary artery catheter. It uses a Peltier effect solid-state cooler and pneumatically driven syringe injector to measure cardiac output once per minute. SETTING: Animal laboratory. ANIMALS: Eight adult Duroc swine weighing between 38.0 and 57.5 kg. INTERVENTIONS: Once each minute, 2 mL of cooled 5% dextrose was injected through the pulmonary catheter. Once every 8 mins, four sequential measurements of cardiac output were made using 10-mL injections. MEASUREMENTS AND MAIN RESULTS: A total of 1249 paired waveforms were processed with both a conventional algorithm and with a neural network. For the conventional algorithm, the correlation coefficient was r2 = .92 and the SD of the difference was 1.30 L/min. For the neural network, the correlation coefficient was r2 = .94 and the SD of the difference was 0.88 L/min. Output filtering improved the results in both cases. CONCLUSION: Neural networks accurately derive cardiac output from 2-mL bolus thermodilution injections, allowing cardiac output to be monitored automatically once per minute in many patients. The technique is convenient and uses standard low-cost catheters.