⚡ electrical engineering

Framework for Patient-Device Model-Based In Silico Clinical Trials in Mechanical Ventilation: A Case Study

This paper presents a standardized, regulatory-aligned framework for conducting in silico clinical trials of mechanical ventilation devices, demonstrating its utility through a case study that systematically evaluates an automated weaning protocol across a diverse virtual patient cohort to identify performance risks and failure modes prior to clinical investigation.

Original authors: Carlotta Hennigs, Georg Männel, Franziska Bilda, Helene Selpien, Dirk Schädler, Michael Imhoff, Philipp Rostalski

Published 2026-10-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Carlotta Hennigs, Georg Männel, Franziska Bilda, Helene Selpien, Dirk Schädler, Michael Imhoff, Philipp Rostalski

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the intensive care units of hospitals around the world, a machine often breathes for patients who cannot do it themselves. This machine, a ventilator, pushes air into the lungs to keep oxygen flowing and carbon dioxide out. For decades, doctors have manually adjusted the settings on these machines, watching monitors and guessing how a specific patient's body will react to a change. But as medicine advances, these machines are becoming smarter. They are being programmed with automated systems that can adjust their own settings in real time, reacting to the patient's changing needs without human intervention. The challenge for engineers and regulators is that these new, self-adjusting systems are incredibly complex. They interact with human biology in ways that are difficult to predict, and testing them on real patients carries significant risks. If a new automated system fails, it could harm a vulnerable person. This creates a need for a safe, virtual way to test these machines against thousands of different body types and disease conditions before they ever touch a human patient.

Researchers at the University of Lübeck and University Hospital Schleswig-Holstein have developed a new framework to solve this problem. They created a method for running "in silico" clinical trials, which are essentially high-fidelity computer simulations of medical experiments. Instead of recruiting real people, they built a digital population of four hundred virtual patients. Each of these digital patients was given a unique set of characteristics, mimicking the wide variety of real people found in intensive care units, from those with stiff, damaged lungs to those with weak breathing muscles. The researchers then connected these virtual patients to a simulated ventilator running an automated weaning protocol. Weaning is the process of helping a patient transition from breathing entirely with a machine to breathing on their own. The goal of the study was to see if this automated system could safely guide patients through this difficult process across a vast range of scenarios, including rare and dangerous edge cases that might not appear in a standard clinical study.

The team did not simply run the simulation once and hope for the best. They designed a rigorous workflow that started by asking specific questions about what the machine needed to do. They wanted to know if the machine could keep the amount of air delivered to the lungs within a safe range, if it could maintain the right breathing rate, and if it could keep carbon dioxide levels from getting too high. They also wanted to see if the machine would ever get out of sync with the patient, causing the patient to fight against the machine, or if the pressure settings would start to oscillate wildly. To test this, they subjected the four hundred virtual patients to nine different scenarios. Some scenarios involved sudden changes in how stiff the lungs were, others simulated a drop in the patient's drive to breathe, and some mimicked severe problems with how the lungs exchange gases. They even included a scenario where the patient's sedation was slowly reduced, mimicking the real-world process of waking a patient up to let them breathe on their own.

The results of these thousands of computer runs revealed both the strengths and the limits of the automated system. In most situations, the machine performed well, keeping the breathing rate and carbon dioxide levels within acceptable targets. It successfully managed the transition for the majority of the virtual patients. However, the simulation also highlighted specific situations where the system struggled. When the virtual patients had severe problems with gas exchange, such as when their lungs were filled with fluid or when a large portion of their blood was bypassing the lungs, the machine had trouble maintaining lung-protective targets. In these difficult cases, the system faced a fundamental trade-off: it could either clear carbon dioxide effectively or keep the pressure in the lungs low enough to prevent injury, but it could not always do both at the same time. The study also found that the machine occasionally fell out of sync with the patient or caused the pressure to oscillate, but these events happened at a frequency that closely matched what doctors observe in real hospitals. Specifically, the simulation showed pressure oscillations in 11.1 percent of the cases, which is very close to the 15.5 percent rate seen in actual clinical data.

Perhaps the most valuable finding was not just how the machine performed, but what the simulation revealed about the factors that influenced that performance. By analyzing the data from the four hundred virtual patients, the researchers identified that the patient's underlying disease state, their level of sedation, and the initial pressure settings were the most important factors determining whether the automated system would succeed. Simple demographics like age or height mattered much less. This insight allows engineers to focus their future improvements on the specific conditions that cause the most trouble. The study also demonstrated that this virtual approach could identify dangerous failure modes without exposing a single real patient to risk. The researchers found that the automated system would likely fail in patients with the most severe forms of lung disease, a finding that suggests these patients might need to be excluded from automated weaning protocols in the real world.

This work represents a significant step forward in how medical devices are tested and regulated. The researchers have shown that it is possible to create a reusable, standardized framework that can stress-test complex, automated medical functions against a wide spectrum of human biology. By using a top-down approach that starts with a clear question and works backward to build the necessary virtual population, they ensure that the results are directly relevant to real-world safety and performance. The study confirms that while these automated systems hold great promise for standardizing care and reducing the workload on doctors, they are not a one-size-fits-all solution. They work well for many patients but require careful monitoring and specific safety limits for those with the most severe conditions. The framework provides a clear path for regulators and manufacturers to verify these systems before they are used in hospitals, ensuring that the technology is robust enough to handle the messy, unpredictable reality of human physiology.

Technical Summary: Framework for Patient-Device Model-Based In Silico Clinical Trials in Mechanical Ventilation

Problem Statement
The development and regulatory evaluation of automated medical devices, particularly those employing multi-variable, adaptive, and AI-driven closed-loop control strategies, face significant challenges. Conventional preclinical and clinical studies are often too expensive, time-consuming, and limited in scope to cover the full spectrum of clinically relevant conditions, rare high-consequence events, and complex patient-device interactions. Furthermore, regulatory frameworks such as the EU Medical Device Regulation (MDR 2017/745) and guidelines from the FDA and ASME (specifically ASME V&V 40) demand rigorous evidence for safety and efficacy, especially for devices moving beyond static, single-variable feedback controllers. Existing in silico clinical trials (ISCTs) for mechanical ventilation often rely on simplified couplings between patient physiology and device logic, omitting critical measurement-chain dynamics and device-specific control algorithms. There is a lack of an integrated, end-to-end pathway that coherently couples respiratory mechanics, gas exchange, respiratory drive, and device logic while explicitly aligning with regulatory credibility frameworks.

Methodology
The authors propose a reusable, device-agnostic ISCT framework grounded in a top-down, risk-informed design logic aligned with ASME V&V 40. The workflow proceeds through eight sequential steps:

Definition of Question of Interest (QoI), Context of Use (COU), and Quantities of Interest (QoI): Establishing the clinical question and the specific role of the model in decision-making.
Model Definition and Development: Creating a Patient-Device Model (PDM) that integrates respiratory mechanics, gas exchange, respiratory control, and device dynamics (including sensor latencies and control logic). The model must satisfy requirements for physiological fidelity, patient-specificity, device coupling, scenario responsiveness, and computational tractability.
Cohort and Scenario Definition: Constructing a virtual cohort that includes both representative patient strata and edge cases (e.g., severe ARDS, deep sedation) to stress-test controller robustness. Scenarios are defined to perturb specific physiological dimensions (mechanics, drive, gas exchange).
Parameter Sampling and Simulation Execution: Using a hybrid sampling strategy (Latin Hypercube Sampling for representative cases and uniform sampling for edge cases) to generate a virtual population.
Result Analysis: Quantifying performance using metrics such as Time in Target (TiT), overshoot, longest excursion time, and composite scores. Patient-ventilator asynchrony and pressure oscillations are detected via specific algorithms.
Answer and Discussion: Synthesizing results to answer the initial QoI, with credibility activities (verification, validation, sensitivity analysis) distributed across all workflow steps rather than confined to a single phase.

This framework is instantiated through a case study evaluating an Automated Weaning Protocol (AWP), inspired by the clinically established SmartCare®/PS system. The study simulated 400 virtual adult ICU patients across 9 perturbation scenarios (including changes in lung elastance, airway resistance, respiratory drive, sedation levels, and gas exchange), resulting in 3,600 total simulation runs.

Key Contributions
The paper identifies three primary contributions:

A Regulatory-Aligned ISCT Workflow: A reusable, top-down framework that enforces traceability from clinical questions to quantitative conclusions, ensuring that model complexity and cohort design are driven by the specific Context of Use and decision risk, consistent with ASME V&V 40.
Virtual Cohort and Scenario Methodology: A specific methodology for generating virtual cohorts that explicitly combines representative and edge-case strata (60%/40% split) using hybrid sampling. This enables systematic stress-testing of controller robustness across the full spectrum of mechanical ventilation pathophysiology, including conditions rarely encountered in conventional trials.
Framework Validation via Case Study: A demonstration of the framework's utility in evaluating an AWP. The study successfully generated decision-relevant evidence, identified performance limits, and reproduced emergent controller behaviors (such as pressure oscillations) without exposing patients to risk.

Results
The case study yielded the following findings:

Performance: The automated weaning protocol generally maintained acceptable working ranges for tidal volume ( $V_T$ ), respiratory rate (RR), and end-tidal carbon dioxide ( $etCO_2$ ) across most scenarios. However, lung-protective targets (specifically plateau pressure limits) were challenged under severe gas exchange impairment (shunt and dead space scenarios).
Asynchrony and Oscillations: Patient-ventilator asynchrony occurred in a phenotype- and scenario-dependent manner. Inspiratory pressure ( $P_{insp}$ ) oscillations were observed in 11.1% of simulations, a frequency comparable to clinical observations (15.5%). These oscillations were primarily associated with ARDS phenotypes.
Influence Analysis: Spearman correlation analysis identified disease state, sedation level, and initial inspiratory pressure as the dominant determinants of AWP performance. Demographic variables (age, height, BMI) showed only weak associations.
Comparison with Clinical Data: The in silico results qualitatively aligned with retrospective clinical data from SmartCare®/PS weaning, reproducing trends in performance difficulty and oscillation rates. Quantitative discrepancies in absolute "Time in Target" values were attributed to the absence of real-world confounders (e.g., leaks, pain, medication changes) in the simulation environment.

Significance and Claims
The authors claim that this framework provides a standardized, credible pathway for the preclinical evaluation of automated ventilator functions. Its significance lies in its ability to:

Support Regulatory Submission: By aligning with ASME V&V 40 and FDA guidance, it offers a structured method for generating evidence that addresses the increased regulatory burden for adaptive and multi-variable devices.
Enable Systematic Stress-Testing: It allows for the identification of failure modes and performance limits (e.g., the trade-off between CO2 clearance and lung protection in severe gas exchange impairment) across broad parameter spaces and edge cases that are under-represented in conventional trials.
Facilitate Iterative Design: The framework supports transparent, decision-relevant evidence generation, enabling iterative refinement of device algorithms prior to clinical investigation.
Reproduce Emergent Behavior: The PDM-based approach successfully captures complex, emergent closed-loop dynamics (like pressure oscillations) that simpler, open-loop simulation frameworks cannot.

The paper modestly positions the case study as a validation of the framework methodology rather than a definitive clinical evaluation of the specific AWP. It acknowledges limitations, such as the exclusion of multi-morbidity and long-term physiological drift, noting that the current implementation serves as a proof-of-concept for a dynamic, extensible methodological platform.

Technical Summary: Framework for Patient-Device Model-Based In Silico Clinical Trials in Mechanical Ventilation

More like this