Anesthesia depth stops relying on population averages: per-patient calibration brings BIS prediction error to 11.77%
Synopsis
Predicting the bispectral index (BIS) during balanced anesthesia with sevoflurane or desflurane alongside propofol and remifentanil has relied on fixed population pharmacodynamic parameters. A computational modelling study merged all three drugs into one response-surface model and fitted per-patient sensitivity parameters from 20 real surgical records using artifact-resistant global optimization. With personalized parameters the cohort mean MDAPE was 11.77%, below the 15% clinical viability threshold; a population baseline aggregated from the same cohort gave 16.98%.
Interpretation
Depth-of-anesthesia prediction moves from one set of population-averaged parameters to sensitivity values fitted separately to each patient's surgical record. Standard pharmacodynamic models treat the maximum hypnotic drop and the synergy steepness as fixed, for example a maximum effect of 70 and a Hill coefficient of 1.2, values derived from controlled volunteer studies. Across 20 balanced-anesthesia patients, the personalized parameters gave a mean MDAPE of 11.77% (median 11.40%, IQR 9.08–14.02%), below the 15% clinical viability threshold.
The fitted sensitivity parameters differ widely between patients, so a single average struggles to cover a real population. Personalized maximum effect Emax ranged from 40.42 to 73.86 and the Hill coefficient γ from 1.29 to 3.40, while literature baselines typically fix γ near 1.2. For the same cohort, a two-stage aggregated population baseline gave a mean MDAPE of 16.98% (median 16.69%, IQR 10.75–22.58%), with a paired Wilcoxon test reporting a mean difference of 5.21 percentage points, 95% CI [2.03, 8.38], p=0.0023.
The model places intravenous and inhaled anesthetics in one pharmacodynamic response surface rather than handling only pairwise combinations. Existing response surfaces that include a volatile agent are mostly two-drug axes derived from healthy volunteers; a unified three-drug surface calibrated directly on real balanced-anesthesia records had not been reported. Normalized propofol and volatile potencies are summed into a hypnotic drive, remifentanil multiplies that synergy, and a sigmoidal Emax equation outputs the predicted BIS.
The optimization suppresses electrical interference and nociceptive spikes during surgery so the curve is not pulled off by artifacts. With a conventional mean-squared-error objective, large artifacts are squared and the optimizer distorts the dose-response curve to chase them. The piecewise Huber loss is quadratic within 5 BIS points of residual, linear from 5 to 25 points, and discards residuals beyond 25 points; the global optimizer runs 15 independent starts with γ constrained between 1.0 and 4.0.
Perspective
The results come from retrospective computational modelling of 20 adult surgical patients receiving balanced anesthesia in VitalDB, excluding patients over 75, those with BMI at least 30 or at most 18.5, and records with severe sparsity; it suits readers following anesthetic pharmacodynamic modelling and closed-loop control who want the current scope of per-patient calibration.
The personalized Emax and γ were fitted and evaluated on the same full intraoperative record for each patient, so MDAPE measures goodness of fit rather than out-of-sample prediction; the authors state that parameters should be calibrated on induction and early maintenance and then evaluated prospectively on late maintenance and emergence. The cohort is only 20 patients, limiting the precision of aggregated population estimates and the generalizability of the parameter ranges; because MDAPE divides by the actual BIS value, equal absolute errors are amplified during deep hypnosis.