Perinatal & Pediatric
Reducing the error in neonatal blood pressure measurements: A calibration study Matthew Loop* Matthew Loop Jieun Park Sarah C Lotspeich Daniel I Feig
Blood pressure is an important vital sign for neonates. Low blood pressures can lead to poor organ perfusion and dangerous intestinal injury, but high blood pressures can also damage the eyes or brain. Despite the potentially life-threatening outcomes of low or high blood pressures, most neonates have their blood pressures measured using an oscillometric cuff, a method known to have large measurement error in routine clinical use. However, a small subset of the sickest neonates have arterial catheters that produce better blood pressure measurements. Some of these neonates also have concurrent cuff-based measurements. Therefore, we conducted a retrospective, EHR-based calibration study of mean arterial pressure (MAP) measurements in a large, academic neonatal intensive care unit. We evaluated linear regression (with and without a spline for cuff-based MAP) and a random forest with 1,000 trees to calibrate the cuff-based MAPs to the catheter-based MAPs, using neonatal and linked maternal covariates. Among 12,966 paired catheter-based and cuff-based MAP measurements across 1,528 neonates, a linear regression model trained on a random 80% subsample: (i) reduced the error of the cuff-based MAPs by 23% from 14.9 mmHg to 11.5 mmHg in the test sample; (ii) produced a calibrated MAP that was closer to the catheter-based MAP than the cuff-based measurement 64% of the time; and (iii) was on average 2.6 mmHg closer to the catheter measurement than the error-prone cuff one. There were also limitations in how well the model calibrated high cuff-based MAP measurements. For catheter-based MAPs above 50 mmHg, the calibration model underestimated these MAPs. (See Figure 1.) Future work will evaluate a broader range of supervised learning methods, identify potentially missing variables that can help improve the underestimation at high MAPs, and transport the calibration model to a population with only cuff-based MAPs available, a population of high clinical importance.
