Virtual patient model for evaluating automated inspired oxygen control

Abstract

Physiologic closed-loop control (PCLC) of inspired oxygen during mechanical ventilation involves frequent adjustment of inspired oxygen fraction (F_iO₂) based on feedback from monitoring peripheral oxygen saturation (S_pO₂). Safety assessments of PCLC algorithms are important prerequisites for patient care, but clinical trials often fail to represent worst-case scenarios that identify limits of safe usage and often do not quantify performance sensitivity to physiologic deviations. The objective of this study was to develop and validate a virtual patient model of pulmonary and systemic gas exchange to assess the safety and efficacy of a PCLC algorithm for F_iO₂ control. A large-scale (3,780,000 simulated patients) virtual observational study was conducted with three clinically relevant scenarios: (1) a step change in patient cardiorespiratory condition; (2) a step change in target S_pO₂; and (3) a step change in PCLC activation. Virtual patients were simulated using a uniform sampling approach to evaluate controller performance in challenging and extreme conditions representing worst-case scenarios. Results in the virtual cohort are not intended to convey predictions of controller performance in typical real-world cohorts. The results demonstrate that PCLC of F_iO₂ is effective for reducing the duration and severity of desaturation during a sudden change in patient condition, and in many cases prevents desaturation altogether (e.g., reducing the occurrence of prolonged desaturation from 69.8 % to 1.5 %). Performance was most sensitive to the physiologic delay between changes in arterial and peripheral oxygenation saturations. Longer physiologic delays (120–300 s) coupled with positive S_pO₂ sensor bias (1.5–3.0 %) were also associated with increased likelihood of system response oscillations. The impact of initial F_iO₂ setting on performance metrics was nonuniform (although 0.4 initial F_iO₂ was optimal in most cases), and was most affected by variations in pulmonary shunt fraction and S_pO₂ sensor bias. This study demonstrates the utility of large-scale virtual patient modeling for sampling wide ranges of physiologic parameters using a multifactorial approach. Sampled conditions may be rarely observed in clinical practice or underrepresented in clinical trials yet warrant careful consideration when evaluating safety and efficacy of autonomous medical device control. The potential impact of the virtual patient model and proposed study design is improved rigor in the evaluation of medical device safety and efficacy, achieved by using computational modeling to complement the shortcomings of clinical trials.