Agreement without accuracy in human endotoxemia: narrow ensemble spread, shared prediction error, and action-library-dependent decisions
Background. Dynamic models of inflammation are increasingly proposed as the basis for closed-loop, automated intervention. Three quantities are used in practice to argue that such a model is fit for that use: whether its parameters are identifiable, how widely a set of plausible fits disperses in its predictions, and whether those fits agree on what to do. Whether any of the three tracks out-of-condition predictive error has rarely been examined on measured human data. Objective. To measure, on published human endotoxemia time courses, how a set of near-optimal fits of one reduced cytokine model behaves when transferred to conditions it was not calibrated on, and whether the dispersion within that set, or the interventions its members recommend, give any warning of the resulting error. Data. Aggregate published cytokine curves from three human experimental endotoxemia studies, redistributed as digitized series in the code archive accompanying Relouw and colleagues (2024): Dillingh and colleagues (2014), LPS boluses of 0.5, 1 and 2 ng/kg; Taudorf and colleagues (2007), a bolus of 0.3 ng/kg; and Kiers and colleagues (2017), boluses of 1 and 2 ng/kg and a combined protocol of a 1 ng/kg bolus followed by 1 ng/kg per hour for three hours, which delivers four times the total LPS of the 1 ng/kg bolus. Methods. An eight-state reduced model of the chain from stimulus through cell activation and mRNA to cytokine turnover, with IL-10 feedback, was calibrated on one cohort by 48 multistart least-squares fits on log-transformed parameters. Four analyses followed: transfer to the six held-out conditions; a decision comparison by exhaustive enumeration over 26 action-and-timing combinations under four objectives and three action libraries; a sampling of the calibration-feasible region by boundary walks in 262 directions; and a falsification test enlarging the model family with a desensitization state. Results. All 48 multistarts converged to calibration errors spanning 0.00025 in RMSE of log10 residuals, while the TNF transcription, mRNA turnover and clearance parameters each spanned a factor of about 17.6 and the IL-10 feedback parameter a factor of 31.9. That last range proved to be a range over inactive values: the maximum transcriptional attenuation attributable to IL-10 feedback was 1.94 percent. On the held-out combined protocol the fits tracked observed TNF to within 7 percent up to 2 hours, then over-predicted by factors of 2.28, 5.19, 5.24 and 3.18 at 3, 4, 6 and 8 hours, while the maximum-to-minimum ratio across the 48 fits stayed between 1.005 and 1.038. Sampling the feasible region did not relieve this: at 4 hours the sampled range ran from 391 to 674 pg/mL against 97.6 observed, and at 6 hours from 211 to 368 against 52.6. Under amplitude-modulating interventions all 48 fits selected the same action under all four objectives; under clearance-modulating interventions one objective split the set 26 against 22, with a maximum relative regret of 0.008. Conclusions. In this case the dispersion of near-optimal fits, and the agreement of the interventions they recommend, co-occurred with a large shared prediction error on a held-out condition, and neither gave warning of it. Two mechanisms in the model were unconstrained by the calibration protocol and contributed at most a few percent of transcriptional damping in every fit found. The held-out protocol differs from the calibration cohort in both schedule and total dose, and the available cohorts cannot separate those two causes. The hypothesis with which this study began, that predictively equivalent fits would recommend materially different interventions, was not observed in the original action library and is reported as not observed. Status. Computational study based entirely on previously published aggregate human data. No new human subjects were recruited, no clinical validation was performed, and no therapeutic claim is made.
Authors
- Piotr Kukier
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24424777
- Primary Topic
- Sepsis Diagnosis and Treatment
- Type
- preprint