Abstract
Background
Mechanical ventilation requires repeated adjustment to changing patient physiology, but consistent individualized management remains challenging. We developed VentPilot, an artificial intelligence–based system for recommending ventilator settings, and evaluated it in multicenter retrospective validation cohorts.
Methods
VentPilot was developed using offline reinforcement learning on high-resolution physiologic and ventilator trajectories from Seoul National University Hospital (SNUH), with a reward combining ventilator-free days and intensivist-derived preference feedback. We evaluated VentPilot in an internal validation cohort from SNUH and an external validation cohort from Mayo Clinic. Fitted Q-evaluation assessed the estimated return of VentPilot relative to observed clinician behavior under the prespecified reward function. In a complementary inverse probability–weighted analysis, clinical outcomes were compared between patients with higher versus lower concordance between observed ventilator settings and VentPilot recommendations.
Results
Among 4296 mechanically ventilated adults, 3002 were included in the derivation cohort, 283 in the internal validation cohort, and 1011 in the external validation cohort. Under the prespecified reward function, fitted Q-evaluation estimated higher returns for VentPilot than for observed clinician behavior in both validation cohorts, with differences in overall reward of 1.6 (95% CI 0.8–2.5) and 1.3 (95% CI 0.6–2.0), respectively. In inverse probability–weighted analyses, higher concordance with VentPilot recommendations was associated with more ventilator-free days within 28 days, with mean differences of 5.0 days (95% CI 2.4–7.6) and 3.1 days (95% CI 1.7–4.6), respectively. Higher concordance was also associated with lower 28-day mortality and shorter ICU length of stay.
Conclusions
In multicenter retrospective validation cohorts, fitted Q-evaluation estimated higher returns for VentPilot than for observed clinician behavior under the prespecified reward function, while greater concordance between observed care and VentPilot recommendations was associated with more favorable clinical outcomes. Further clinical evaluation is warranted to establish VentPilot’s safety, usability, and clinical impact.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1186/s40560-026-00947-7.
Keywords: Mechanical ventilation, Artificial intelligence, Reinforcement learning, Critical care, Ventilator-free days, Clinical recommendation system
Background
Invasive mechanical ventilation is a cornerstone of supportive care for respiratory failure in intensive care units (ICUs) and is used in more than 20 million patients worldwide each year, supporting gas exchange while allowing time for recovery [1]. However, suboptimal mechanical ventilation can lead to substantial morbidity, including ventilator-induced lung injury, diaphragm dysfunction, and increased mortality [2]. Moreover, extended use of analgesia and sedation to facilitate mechanical ventilation can contribute to immobilization and ICU-acquired weakness, further worsening patient outcomes [3].
Selecting and maintaining appropriate ventilator settings is essential for minimizing lung and diaphragm injuries and improving patient outcomes. The current best practices emphasize lung- and diaphragm-protective ventilation strategies, as well as targeting light sedation [4, 5]. However, consistent implementation of these strategies is challenging because a one-size-fits-all approach does not account for patient heterogeneity and clinical practices vary substantially across institutions [6, 7]. Moreover, a patient’s respiratory mechanics evolve dynamically over time, often necessitating frequent reassessment and adjustment of ventilator settings to maintain appropriate support [8, 9]. Furthermore, there is a limited number of adequately trained clinicians who can consistently match a patient’s respiratory physiology to the appropriate mechanical ventilation parameters in a timely manner throughout the day. While automated systems have been proposed to bridge this gap, many rely on static, rule-based logic that may not adequately adapt to the dynamic and heterogeneous physiology of critically ill patients [10–15]. A recent international randomized clinical trial of automated closed-loop ventilation also showed that, despite improved ventilation quality, ventilator-free days were not significantly increased compared with protocolized conventional ventilation [16].
To address these limitations, we developed VentPilot, an artificial intelligence (AI) system designed to provide real-time, personalized ventilator setting recommendations. Unlike previous approaches, VentPilot leverages a prospectively collected, large-scale, high-resolution dataset and incorporates expert clinical judgment through reinforcement learning from human feedback (RLHF) [17, 18]. This study aimed to evaluate VentPilot’s ventilator-setting recommendation policy and the association between concordance with its recommendations and clinical outcomes, using datasets from Seoul National University Hospital (SNUH) and the Mayo Clinic.
Methods
Study design and population
We conducted an international, multicenter retrospective analysis using individual ICU patient data. For model derivation and internal validation, we used a prospective database from SNUH, a 1787-bed tertiary referral hospital in South Korea; for external validation, we used a retrospective database from Mayo Clinic, a 2059-bed quaternary referral hospital in the United States. The study included adult patients (> 17 years) who received invasive mechanical ventilation in medical or surgical ICUs between October 2019 and December 2023 at SNUH, and between February 2017 and May 2023 at the Mayo Clinic. Exclusion criteria for the derivation dataset included (1) patients who received extracorporeal membrane oxygenation (ECMO) and (2) those lacking data on ICU admission type, intubation/extubation dates, vital signs, arterial blood gases, or ventilatory parameters throughout the intubation period. For validation, patients were additionally required to have received invasive mechanical ventilation for at least 36 h to ensure sufficient longitudinal follow-up for policy evaluation and concordance assessment. Additional exclusion criteria are provided in Additional file 1: Methods S1.
To construct a new high-resolution dataset for mechanical ventilation, we prospectively collected patient data from SNUH ICUs at a two-second sampling rate by installing specialized recording devices on the ventilators and hemodynamic monitoring systems at each ICU bed [19]. At the Mayo Clinic, we accessed retrospective ICU data with an average sampling rate of fifteen minutes through the Mayo Clinic Platform_Accelerate program, querying their Electronic Health Record (EHR) system via Spark SQL.
This study was conducted in compliance with ethical guidelines, including the revised Helsinki Declaration and ICH-GCP standards. The Institutional Review Board of SNUH approved the study (approval number H-2109-143-1258). Given that the study was observational and patient identifiers were removed, the requirement for written informed consent was waived. This study analyzes de-identified EHR data via Mayo Clinic Platform Discover. The data presented here is extracted following a privacy-preserving protocol and deemed de-identified by an expert in accordance with HIPAA Privacy Rule. Additional data, including raw EHR data, cannot be shared due to expert determination. Contact corresponding authors for more information on Mayo Clinic Platform Discover.
Data preparation
At SNUH, data were allocated to derivation and internal validation cohorts using a temporal split (Additional file 1: Fig. S1). Patients admitted before October 2022 comprised the derivation cohort, whereas those admitted in or after October 2022 were assessed for validation eligibility and included in the internal validation cohort if they met the prespecified criteria.
Each patient’s raw data were transformed into a trajectory, defined as a multivariable time series of physiologic measurements, electronic medical record variables, and ventilator parameters spanning the intubation period. Additional details on the trajectory construction are provided in Additional file 1: Methods S2 and Table S1.
Model training
VentPilot was developed using reinforcement learning, a machine learning approach in which a model learns a decision policy that links the current patient state to actions that maximize cumulative reward over time (Fig. 1). The action space comprised six ventilator control variables routinely adjusted during mechanical ventilation: set fraction of inspired oxygen (FiO₂), set positive end-expiratory pressure (PEEP), set pressure control/support above PEEP (for pressure-targeted modes), set tidal volume (for volume-targeted modes), and set respiratory rate. All remaining clinical and ventilator variables were used as state inputs. The reward was defined as a weighted sum of a ventilator-free days (VFDs)–based component and a preference-based component derived from a learned preference model trained on intensivist-reviewed state-action pairs. The model was then trained on the derivation cohort to maximize cumulative reward using an offline reinforcement learning framework, with hyperparameters selected by 5-fold cross-validation [20, 21]. Additional details on the model training procedure are provided in Additional file 1: Methods S2–S5 and Figure S2.
Fig. 1.
Overview of the VentPilot development and validation pipeline. High-resolution physiologic and ventilator data were prospectively collected at Seoul National University Hospital (SNUH; n = 4121) and used to construct patient trajectories for model derivation (n = 3002) and internal validation (n = 283). VentPilot development combined (i) expert feedback collection, in which a committee of board-certified intensivists reviewed discrepancy (outlier) events identified by a behavior-cloning model and graded the appropriateness of clinician actions, and (ii) offline reinforcement learning using logged clinician interactions to refine the recommendation policy. VentPilot was evaluated using two complementary approaches: off-policy evaluation with fitted Q-evaluation to estimate expected return under the predefined reward, and inverse probability–weighted analyses comparing clinical outcomes between patients with higher versus lower agreement between observed ventilator settings and VentPilot recommendations. External validation was performed in an independent cohort from the Mayo Clinic (n = 5634; n = 1011 included after applying exclusion criteria)
Model evaluation and statistical analysis
We performed two complementary analyses in the internal and external validation cohorts: off-policy evaluation and inverse probability–weighted outcome analysis. Off-policy evaluation used fitted Q-evaluation to estimate the expected return of VentPilot relative to the observed clinician behavior policy [22]. Policy-value differences were estimated separately in each validation cohort, and trajectory-level nonparametric bootstrapping was used to obtain 95% confidence intervals.
For inverse probability weighting, each patient was assigned an agreement score reflecting the concordance between VentPilot recommendations and observed ventilator settings; details are provided in Additional file 1: Methods S6. Within each validation cohort, patients were classified as AI-similar (top 50% by agreement score) or AI-dissimilar (remaining 50%). Propensity scores for AI similarity were estimated using prespecified baseline covariates, and inverse probability weights targeting the average treatment effect were applied. The primary outcome was VFDs within 28 days, defined as days alive and free from invasive mechanical ventilation during the first 28 days after intubation [23].
Additional descriptive analyses compared observed ventilator settings with VentPilot recommendations across prespecified physiologic strata, evaluated the frequency and timing of VentPilot recommendations and clinician setting adjustments during predefined physiologic conditions, and assessed temporal consistency of recommendation timing. In an additional exploratory comparator analysis, we compared VentPilot with adaptive support ventilation (ASV), an existing rule-based automated ventilation strategy, using the same off-policy evaluation framework as in the primary evaluation and inverse probability–weighted outcome analysis between AI-similar non-ASV care and ASV-exposed care [24]. Details on the descriptive analyses are provided in Additional file 1: Methods S7–S10.
Subgroup analyses were conducted across age, sex, and baseline P/F ratio in both validation cohorts and, in the Mayo Clinic cohort, across selected diagnosis- and comorbidity-defined strata. These analyses were exploratory, and no adjustment for multiple comparisons was performed. Sensitivity analyses for off-policy evaluation included alternative time grids and weighted importance sampling as an alternative estimator [22]. We also varied the preference-reward weight and policy-extraction change-to-no-change sampling ratio and repeated both off-policy and inverse probability–weighted analyses. For the inverse probability–weighted analyses, we additionally reassessed concordance on a fixed 15-min grid, performed a fixed early-window analysis in which baseline covariates were derived from the first 12 h after intubation and concordance was assessed during the subsequent 12–36-h period, applied alternative AI-similarity thresholds, and, in the Mayo Clinic cohort, recalculated concordance using only pressure-targeted ventilation periods.
All statistical tests were two-tailed, and a P value < 0.05 was considered statistically significant. Analyses were performed using R version 4.3.1 and Python version 3.11.5.
Results
Study participants
Of 4121 mechanically ventilated patients screened at SNUH, 3,002 admitted before October 2022 comprised the derivation cohort. Of 1119 admitted from October 2022 onward, 836 were ineligible for validation, most commonly because invasive mechanical ventilation lasted less than 36 h, and 283 formed the internal validation cohort (Additional file 1: Fig. S1). For external validation, 1011 of 5634 screened patients at the Mayo Clinic met eligibility criteria; the most common reason for exclusion was also invasive mechanical ventilation for less than 36 h. Baseline characteristics are shown in Table 1 and Additional file 1: Table S4. For the inverse probability–weighted analysis, standardized mean differences for baseline characteristics were < 0.1 after weighting in both validation cohorts (Additional file 1: Tables S5 and S6).
Table 1.
Baseline clinical characteristics of derivation and validation cohorts
| Variable | SNUH derivation cohort (n = 3002) | SNUH internal validation cohort (n = 283) | Mayo Clinic external validation cohort (n = 1011) | Overall (n = 4296) |
|---|---|---|---|---|
| Type of ICU admission | ||||
| Medical | 452 (15.1) | 135 (47.7) | 600 (59.3) | 1187 (27.6) |
| Surgical | 2550 (84.9) | 148 (52.3) | 411 (40.7) | 3109 (72.4) |
| Age, y | 59.9 ± 15.8 | 67.2 ± 14.2 | 63.2 ± 15.1 | 61.2 ± 15.7 |
| Sex | ||||
| Male | 1654 (55.1) | 177 (62.5) | 566 (56.0) | 2397 (55.8) |
| Female | 1348 (44.9) | 106 (37.5) | 445 (44.0) | 1899 (44.2) |
| Body mass index, kg/m2 | 23.8 ± 4.3 | 22.9 ± 4.3 | 30.4 ± 7.7 | 25.3 ± 6.0 |
| Primary diagnosis | ||||
| Cardiovascular | 554 (18.5) | 21 (7.4) | 220 (21.8) | 795 (18.5) |
| Gastrointestinal | 105 (3.5) | 17 (6.0) | 76 (7.5) | 198 (4.6) |
| Hepatobiliary | 123 (4.1) | 13 (4.6) | 50 (4.9) | 186 (4.3) |
| Neoplastic | 1552 (51.7) | 97 (34.3) | 24 (2.4) | 1673 (38.9) |
| Neurologic | 140 (4.7) | 45 (15.9) | 90 (8.9) | 275 (6.4) |
| Renal | 14 (0.5) | 8 (2.8) | 29 (2.9) | 51 (1.2) |
| Respiratory | 85 (2.8) | 14 (4.9) | 131 (13.0) | 230 (5.4) |
| Sepsis, including pneumonia | 96 (3.2) | 28 (9.9) | 149 (14.7) | 273 (6.3) |
| Other | 333 (11.1) | 40 (14.1) | 242 (23.9) | 615 (14.3) |
| Comorbidities | ||||
| Cardiovascular disease | 745 (24.8) | 108 (38.2) | 725 (71.7) | 1578 (36.7) |
| Diabetes | 487 (16.2) | 79 (27.9) | 346 (34.2) | 912 (21.2) |
| COPD | 129 (4.3) | 28 (9.9) | 245 (24.2) | 402 (9.4) |
| Other chronic lung diseasea | 288 (9.6) | 69 (24.4) | 182 (18.0) | 539 (12.5) |
| Chronic kidney disease | 164 (5.5) | 30 (10.6) | 361 (35.7) | 555 (12.9) |
| Malignancy | 2039 (67.9) | 185 (65.4) | 280 (27.7) | 2504 (58.3) |
| P/F ratio, mmHg | 383 (276–488) | 268 (196–387) | 247 (175–332) | 328 (221–440) |
| SOFA score | 8 (6–12) | 12 (9–15) | 10 (8–12) | 10 (7–12) |
| APACHE II score | 25 (21–29) | 31 (26–35) | 25 (22–28) | 25 (22–29) |
APACHE Acute Physiology and Chronic Health Evaluation, COPD chronic obstructive pulmonary disease, ICU intensive care unit, SNUH Seoul National University Hospital, SOFA Sequential Organ Failure Assessment
Values are reported as n (%) for categorical variables and median (IQR) or mean ± SD for continuous variables
a Includes interstitial lung disease or other chronic lung diseases, excluding COPD
Off-policy evaluation
Fitted Q-evaluation estimated a higher return for VentPilot than for the observed clinician behavior policy in both the SNUH internal validation cohort (difference 1.6; 95% CI 0.8–2.5) and the Mayo external validation cohort (difference 1.3; 95% CI 0.6–2.0) (Table 2 and Additional file 1: Table S7). Similar findings were observed across both reward components (VFDs reward and preference reward). In ablation analyses, removing intensivist feedback (i.e., training without the preference reward) and training without high-resolution training trajectories both attenuated the estimated return difference. In an additional calibration analysis, FQE estimates of clinician behavior were also close to empirical clinician returns in both cohorts (Additional file 1: Table S14).
Table 2.
Fitted Q-evaluation of VentPilot and ablated policies in validation cohorts
| Reward type | AI | Without intensivist feedback | Without high-resolution train data | Without intensivist feedback and high-resolution train data |
|---|---|---|---|---|
| SNUH internal validation cohort | ||||
| Overall | 1.6 (0.8–2.5) | 1.2 (0.5–1.8) | 1.3 (0.5–1.7) | 0.9 (0.2–1.4) |
| VFDs, daysa | 1.5 (0.8–2.9) | 1.0 (0.2–2.1) | 1.1 (0.4–2.1) | 0.8 (0.1–1.8) |
| Preference | 1.5 (0.1–2.8) | -0.2 (− 0.7–0.8) | 1.2 (− 0.4–2.7) | -0.5 (− 1.1–0.7) |
| Mayo Clinic external validation cohort | ||||
| Overall | 1.3 (0.6–2.0) | 0.7 (0.0–1.5) | 0.8 (0.1–1.5) | 0.2 (− 0.7–1.0) |
| VFDs, daysa | 1.2 (0.3–2.2) | 1.0 (0.2–1.9) | 0.6 (− 0.2–1.4) | 0.1 (− 0.7–1.0) |
| Preference | 0.8 (0.5–1.3) | 0.4 (0.0–0.9) | 0.5 (0.1–0.8) | 0.2 (− 0.2–0.5) |
AI artificial intelligence, SNUH Seoul National University Hospital, VFDs ventilator-free days
Values are differences in estimated return of the AI policy minus the return of the clinician policy; positive values indicate higher estimated return than the clinician policy. Overall reward is the weighted sum of a VFDs-based reward and a preference reward generated by a learned preference model trained on intensivist action ratings. Fitted Q-evaluation was performed separately for the overall reward and for each reward component using the corresponding reward definition. “without intensivist feedback” denotes training with the preference reward weight set to 0 (i.e., no preference-based supervision signal). “without high-resolution train data” denotes training without the prospectively collected high-resolution time-series dataset, using only the coarser electronic medical record-derived dataset available in routine documentation
a For the VFDs component, reported differences can be interpreted as estimated differences in VFDs (in days), because the VFD-based return differs from VFDs only by a constant shift
Inverse probability–weighted analysis
After inverse probability weighting, the mean number of VFDs was higher in the AI-similar group than in the AI-dissimilar group in both validation cohorts (Table 3). In the SNUH internal validation cohort, mean VFDs were 17.2 ± 10.5 days in the AI-similar group compared to 12.2 ± 11.5 days in the AI-dissimilar group, with a mean difference of 5.0 days (95% CI 2.4–7.6; P < 0.001). In the external validation cohort, the mean number of VFDs was significantly higher in the AI-similar group at 17.9 ± 10.9 days compared to 14.9 ± 11.8 days in the AI-dissimilar group, with a mean difference of 3.1 days (95% CI 1.7–4.6; P < 0.001).
Table 3.
Primary and secondary outcomes of the AI-similar and AI-dissimilar groups after inverse probability weighting
| Outcome | SNUH internal validation cohort | Mayo Clinic external validation cohort | ||||
|---|---|---|---|---|---|---|
| AI-similar (n = 137)a | AI-dissimilar (n = 138)a | Absolute difference (95% CI)b | AI-similar (n = 512)a | AI-dissimilar (n = 505)a | Absolute difference (95% CI)b | |
| Primary outcome—ventilator-free days (within 28 days) | ||||||
| Mean | 17.2 ± 10.5 | 12.2 ± 11.5 | 5.0 (2.4–7.6) | 17.9 ± 10.9 | 14.9 ± 11.8 | 3.1 (1.7–4.6) |
| Median | 22.8 (0.0–25.0) | 16.6 (0.0–23.5) | 6.1 (1.9–23.0) | 24.0 (0.0–25.4) | 22.3 (0.0–25.1) | 1.7 (0.8–2.9) |
| Secondary outcomes | ||||||
| Ventilation duration among ICU survivors, days | 3.6 (2.2–5.8) | 5.7 (3.7–8.2) | − 2.1 (− 2.8 to − 0.9) | 2.7 (1.9–4.2) | 3.9 (3.0–5.5) | − 1.2 (− 1.5 to − 0.9) |
| Ventilation duration among in-hospital survivors, days | 3.6 (2.2–5.8) | 5.5 (3.5–7.6) | − 1.9 (− 2.7 to − 0.3) | 2.6 (1.8–4.2) | 3.9 (3.0–5.4) | − 1.2 (− 1.5 to − 0.8) |
| 28-day mortality, n (%) | 39 (28.5) | 58 (41.7) | − 13.2 (− 21.6 to − 4.7) | 134 (26.2) | 190 (37.7) | − 11.4 (− 17.1 to − 5.7) |
| ICU mortality, n (%) | 28 (20.3) | 47 (34.1) | − 13.8 (− 21.2 to − 6.3) | 86 (16.7) | 125 (24.8) | − 8.0 (− 13.0 to − 3.1) |
| In-hospital mortality, n (%) | 51 (37.4) | 66 (47.7) | − 10.3 (− 23.2 to − 1.2) | 139 (27.1) | 185 (36.7) | − 9.6 (− 15.3 to − 3.9) |
| ICU length of stay, days | ||||||
| Total | 6.9 (4.3–10.1) | 8.5 (6.0–11.5) | − 1.6 (− 3.3 to − 0.2) | 4.7 (3.4–6.8) | 6.0 (4.3 to 8.3) | − 1.3 (− 1.7 to − 0.9) |
| ICU survivors | 6.9 (4.4–10.3) | 8.5 (6.7–11.3) | − 1.6 (− 3.3 to − 0.2) | 4.9 (3.4–7.2) | 6.0 (4.3 to 8.0) | − 1.1 (− 1.5 to − 0.6) |
| ICU non-survivors | 7.0 (3.5–9.6) | 8.1 (3.8–14.5) | − 1.1 (− 5.8–1.7) | 4.4 (2.8–5.9) | 6.2 (4.4 to 9.2) | − 1.9 (− 2.9 to − 1.1) |
| Hospital length of stay, days | ||||||
| Total | 20.3 (9.2–42.1) | 20.9 (12.0–40.6) | − 0.6 (− 6.8–5.9) | 14.6 (8.4–26.2) | 13.3 (8.2–22.7) | 1.3 (− 0.7–3.0) |
| Hospital survivors | 27.7 (15.4–49.4) | 38.6 (24.5–61.1) | − 10.9 (− 20.7–0.9) | 16.2 (10.8–28.7) | 16.9 (11.0–32.2) | − 0.7 (− 2.4–2.4) |
| Hospital non-survivors | 13.3 (3.6–28.5) | 12.6 (8.1–20.0) | 0.8 (− 7.6–5.9) | 6.4 (4.0–13.2) | 8.4 (5.4–13.5) | -2.0 (− 3.5 to − 0.1) |
AI artificial intelligence, CI confidence interval, ICU intensive care unit, SNUH Seoul National University Hospital
Values after weighting are presented as weighted n (%) for binary outcomes and weighted mean ± SD or weighted median (IQR) for continuous outcomes. The primary outcome is presented as both weighted mean ± SD and weighted median (IQR); other continuous outcomes are presented as weighted median (IQR)
a The n values indicate weighted sample sizes, calculated as the sum of the final stabilized inverse probability weights after truncation at the 99th percentile and rounded to the nearest integer
b For continuous outcomes, absolute differences are in days. For binary outcomes, absolute differences are risk differences expressed in percentage points
Compared with the AI-dissimilar group, the AI-similar group had fewer days of mechanical ventilation among ICU survivors in both the SNUH internal validation cohort (median, 3.6 vs. 5.7 days; difference, −2.1 days; 95% CI −2.8 to −0.9; P < 0.001) and the Mayo external validation cohort (median, 2.7 vs. 3.9 days; difference, −1.2 days; 95% CI −1.5 to −0.9; P < 0.001). The AI-similar group also had lower 28-day mortality in both cohorts (SNUH: 28.5% vs. 41.7%; absolute difference, −13.2 percentage points; 95% CI −21.6 to −4.7; P = 0.004; Mayo: 26.2% vs. 37.7%; absolute difference, −11.4 percentage points; 95% CI −17.1 to −5.7; P = 0.006) and shorter ICU length of stay (SNUH: median 6.9 vs. 8.5 days; difference, −1.6 days; 95% CI −3.3 to −0.2; P = 0.01; Mayo: median 4.7 vs. 6.0 days; difference, −1.3 days; 95% CI −1.7 to −0.9; P < 0.001).
Policy behavior and robustness analyses
In descriptive comparisons across prespecified physiologic strata in the Mayo Clinic external validation cohort, the clearest differences were higher sFiO2 recommendations when SpO2 was < 92% (Δ = + 7.1%, d = + 0.31), higher sPEEP when P/F ratio was <100 mmHg (Δ = + 1.2 cmH2O, d = + 0.32), and lower sPEEP when P/F ratio was ≥ 300 mmHg (Δ = − 0.8 cmH2O, d = − 0.28). During assisted ventilation, VentPilot also tended to recommend higher sPS when pH was < 7.3 or TV/IBW was <4 mL/kg, and lower sPC or sTV when TV/IBW was ≥8 mL/kg. Similar directional patterns were observed in the SNUH internal validation cohort (Additional file 1: Figs. S3 and S4).
Across episodes meeting the predefined physiologic conditions, a VentPilot recommendation in the condition-specific direction occurred in 24.1%–93.2% of episodes at SNUH and 11.4%–99.5% at Mayo Clinic, whereas a clinician adjustment in the same direction occurred in 8.1%–38.7% and 7.0%–27.4% of episodes, respectively (all P < 0.001). Among episodes in which both events occurred, the median time from condition onset to the first VentPilot recommendation was 0.0 h across condition types, whereas the median time to the first clinician adjustment in the same direction ranged from 0.1 to 1.0 h at SNUH and from 0.5 to 7.0 h at Mayo Clinic (all P ≤ 0.004) (Additional file 1: Fig. S5 and Table S8).
In the SNUH internal validation cohort, VentPilot recommendation events exhibited less temporal variation across time of day and day of week than clinician setting-adjustment events, with clinician adjustments occurring less often than expected overnight (00:00–05:59, 54% of expected) and on Sunday (78% of expected), but more often during daytime, especially 06:00–11:59 and 12:00–17:59 (128% and 125%, respectively), and on Monday (113% of expected); in contrast, VentPilot recommendation events remained close to the uniform 100% reference across bins (Additional file 1: Fig. S6).
In the exploratory ASV comparator analysis in the Mayo Clinic external validation cohort, VentPilot was evaluated relative to ASV-exposed care. In fitted Q-evaluation among ASV-exposed patients, VentPilot had a higher estimated return on the overall reward than the observed clinician behavior policy under ASV (difference, 1.9; 95% CI 0.2–2.9) (Additional file 1: Table S9). In inverse probability–weighted comparisons between AI-similar non-ASV care and ASV-exposed care, baseline covariates were well balanced after weighting (all standardized mean differences < 0.1; Additional file 1: Table S10). The AI-similar non-ASV group had more VFDs, shorter ventilation duration among ICU survivors, shorter ICU length of stay, and lower 28-day mortality (Additional file 1: Table S11).
In exploratory subgroup analyses, the direction of the findings was generally consistent across age, sex, and baseline P/F ratio strata in both validation cohorts. Among patients with a baseline P/F ratio less than 200, the differences in fitted Q-evaluation were 2.2 (95% CI 0.5–3.9) in SNUH and 1.7 (95% CI 1.2–2.1) in Mayo, with corresponding differences in mean VFDs of 6.7 (95% CI 1.3–11.4; P = 0.01) and 4.5 (95% CI 2.1–6.2; P < 0.001) days, respectively. Findings were also directionally similar in the selected diagnosis- and comorbidity-defined subgroups in the Mayo Clinic cohort (Additional file 1: Table S12).
In the fixed early-window analysis using baseline covariates from 0–12 h and concordance from 12–36 h after intubation, the weighted mean VFD difference between AI-similar and AI-dissimilar patients was 4.1 days (95% CI 1.0–6.7) in the SNUH cohort and 2.8 days (95% CI 1.1–4.2) in the Mayo Clinic cohort. Other sensitivity analyses using alternative time grids, weighted importance sampling, agreement definitions, AI-similarity thresholds, preference-reward weights, policy-extraction sampling ratios, and concordance restricted to pressure-targeted ventilation periods showed directionally similar findings (Additional file 1: Table S13). An illustrative pair of patients with similar baseline propensity scores but different concordance with VentPilot recommendations is shown in Additional file 1: Fig. S7.
Discussion
In this multicenter retrospective study, fitted Q-evaluation estimated higher returns for VentPilot than for observed clinician behavior under the prespecified reward function in both validation cohorts. In complementary inverse probability–weighted analyses, greater concordance between observed ventilator settings and VentPilot recommendations was associated with more VFDs, lower 28-day mortality, and shorter ICU length of stay in both cohorts. Taken together, these findings suggest that VentPilot learned ventilator-setting patterns that were consistently associated with favorable outcomes across both validation cohorts.
While prior studies have explored AI for mechanical ventilation, this study differs from prior work in several aspects [25–27]. Previous work often relied on public ICU datasets with low time resolution (≥ 1 h) or lacked important ventilator variables such as minute ventilation [28–31]. In contrast, we prospectively collected a high-frequency observational dataset focused on mechanical ventilation at SNUH (device-level recordings at 2-s resolution), enabling detailed characterization of ventilatory and physiologic trajectories. Consistent with this design, training ablations showed attenuation when VentPilot was trained without high-resolution time-series information, suggesting that higher-fidelity physiologic trajectories contribute to effective policy learning.
A second distinction was the reward design. Conventional reinforcement learning relies on custom-designed reward functions, which may inaccurately represent desired outcomes and lead to “reward hacking,” where a model achieves a high reward by exploiting these inaccuracies without corresponding real-world benefit [32, 33]. Instead, VentPilot directly anchored the reward to VFDs as a patient-centered outcome and supplemented it with intensivist-derived preference feedback. Established physiologic safety targets were enforced separately through prespecified guardrails rather than encoded as additional reward terms, avoiding the need to assign weights across potentially competing surrogate objectives. Physiologic objectives beyond these guardrails, such as driving pressure, mechanical power, and hypercapnia, were not explicitly optimized. Consistent with the intended role of expert feedback, the estimated return difference decreased when intensivist feedback was removed.
Our findings are also consistent with the broader shift toward personalized mechanical ventilation, which emphasizes adjusting ventilator settings to evolving gas exchange, respiratory mechanics, and patient effort rather than relying on fixed, one-size-fits-all strategies [4–6]. Although closed-loop ventilation systems can improve the selection of lung-protective settings and reduce staff workload, evidence for patient-centered benefits remains mixed [16, 34, 35]. In this context, our exploratory comparison with ASV-exposed care provided context for the broader question of whether the clinical value of automated ventilation depends not only on automation itself, but also on the adaptability and clinical alignment of the underlying decision policy. However, the ASV comparison was exploratory and observational; despite covariate balancing, residual selection bias related to patient characteristics, clinician preference, or institutional practice cannot be excluded.
Comparisons across clinically relevant physiologic strata provide an interpretable assessment of how VentPilot recommendations differed from observed clinician settings. VentPilot generally recommended higher oxygenation support when hypoxemia was more severe, lower PEEP when oxygenation was preserved, greater ventilatory assistance when acidemia and very low tidal volumes suggested possible under-assistance, and less inspiratory pressure or lower set tidal volume when tidal volumes were high. These patterns align with current principles of ventilator management: escalating oxygenation support during hypoxemia, avoiding unnecessary support when oxygenation is preserved, individualizing PEEP according to oxygenation severity and hemodynamic tolerance, and titrating assistance to avoid both insufficient and excessive respiratory effort. Together, these findings support the physiologic coherence of VentPilot’s recommendations [4, 5, 36–41].
Across episodes meeting the predefined physiologic conditions, VentPilot recommendations in the condition-specific direction occurred more frequently and with shorter times from condition onset than clinician adjustments in the same direction. VentPilot recommendation timing also varied less by time of day and day of week than clinician setting changes. Because VentPilot was evaluated automatically at each evaluation timestamp whereas clinician actions occurred asynchronously, these comparisons characterize recommendation availability and directional consistency rather than greater clinical responsiveness or appropriateness. Clinicians may also reasonably maintain current settings based on clinical information unavailable to this analysis or respond through other interventions. Nevertheless, the more uniform temporal distribution of VentPilot recommendations highlights a potential role for continuous and consistent reassessment in ventilator management, which may otherwise be influenced by workflow, staffing patterns, and competing clinical demands [7]. Such a system could complement, rather than replace, bedside judgment by prompting reconsideration of ventilator settings as patient physiology evolves [1].
This study has several limitations. First, off-policy evaluation estimates the performance of VentPilot from historical state-action transitions rather than from actual outcomes observed after applying its recommendations or from explicit modeling of the physiologic consequences of those recommendations. Although VentPilot recommendations broadly overlapped with observed setting distributions across prespecified physiologic strata, limited support in some state-action regions may still increase uncertainty in FQE estimates through Q-function extrapolation. Thus, fitted Q-evaluation provides an estimate of expected return rather than direct evidence of clinical benefit. Second, although inverse probability weighting balanced prespecified baseline covariates, residual time-varying confounding remains possible as concordance was assessed longitudinally over the subsequent ventilation course. To reduce the influence of later clinical course on concordance classification, we performed a sensitivity analysis restricting concordance assessment to a fixed 24-h early window (12–36 h after intubation); the VFD difference was modestly attenuated but remained statistically significant. However, this sensitivity analysis cannot fully eliminate time-varying confounding.
Third, our analysis focused on several key ventilator control variables, excluding other factors such as ventilator mode changes, sedative use, explicit weaning status, and extubation timing. Although spontaneous breathing was partially reflected by ventilator mode and respiratory variables, these broader clinical factors were not explicitly modeled; incorporating them in future research could provide a more complete representation of ventilator management. Intermittently measured laboratory values may also become less representative of current physiology as time since measurement increases, potentially affecting state representation and recommendations despite the use of time-limited carry-forward and measurement-recency information. In addition, although the control-specific trees were fitted using advantage estimates from the joint multidimensional action, combining separate trees may not fully preserve dependencies among simultaneously adjusted ventilator settings. For volume-targeted modes in the Mayo Clinic cohort, sTV recommendations were derived from sPC recommendations using dynamic compliance rather than learned directly from volume-targeted data, which may not fully capture time-varying pressure–volume relationships. The WIS sensitivity analysis similarly factorized target- and behavior-policy probabilities across setting dimensions to avoid sparse estimation of the high-dimensional joint action distribution [42, 43]; this approximation may not fully capture dependencies among simultaneous setting changes.
Fourth, the preference model was intentionally trained on disagreement-enriched state-action pairs and was not independently validated on a separately labeled dataset; therefore, the generalizability of the preference signal across the broader range of ventilator-management decisions remains to be established. Fifth, validation was restricted to patients receiving invasive mechanical ventilation for at least 36 h, which ensured sufficient longitudinal observation but excluded patients with early extubation or early death; similar eligibility approaches have been used in trials of automated ventilation [16]. Exclusion of patients with substantially incomplete trajectory data may also limit generalizability to settings with greater data missingness. Lastly, this observational study did not directly assess the safety, usability, or clinical impact of applying VentPilot recommendations in clinical practice.
Conclusions
Fitted Q-evaluation estimated higher returns for VentPilot than for observed clinician behavior under the prespecified reward function, and closer concordance between observed care and VentPilot recommendations was associated with more favorable outcomes across two validation cohorts. These findings support further clinical evaluation to establish VentPilot’s safety, usability, and clinical impact.
Supplementary Information
Acknowledgements
We thank the patients whose de-identified data contributed to this study. We also thank the clinical, respiratory therapy, nursing and information technology teams at Seoul National University Hospital and Mayo Clinic for their support of data collection, data governance and study operations. We thank the Mayo Clinic Platform Discover team for facilitating privacy-preserving access to the external validation data.
Abbreviations
- AI
Artificial intelligence
- APACHE II
Acute Physiology and Chronic Health Evaluation II
- ASV
Adaptive support ventilation
- CI
Confidence interval
- COPD
Chronic obstructive pulmonary disease
- ECMO
Extracorporeal membrane oxygenation
- EHR
Electronic health record
- FiO2
Fraction of inspired oxygen
- HIPAA
Health Insurance Portability and Accountability Act
- ICH-GCP
International Council for Harmonisation Good Clinical Practice
- ICU
Intensive care unit
- IQR
Interquartile range
- PaCO2
Arterial partial pressure of carbon dioxide
- PaO2
Arterial partial pressure of oxygen
- PEEP
Positive end-expiratory pressure
- P/F ratio
Ratio of PaO₂ to FiO₂
- RLHF
Reinforcement learning from human feedback
- SD
Standard deviation
- sFiO2
Set fraction of inspired oxygen
- sPC
Set pressure control above PEEP
- sPEEP
Set positive end-expiratory pressure
- sPS
Set pressure support above PEEP
- sRR
Set respiratory rate
- sTV
Set tidal volume
- SNUH
Seoul National University Hospital
- SOFA
Sequential Organ Failure Assessment
- SpO2
Peripheral oxygen saturation
- SQL
Structured Query Language
- TV/IBW
Tidal volume normalized to ideal body weight
- VFDs
Ventilator-free days
Author contributions
H.Y.L., G.A., Y.J., S.-M.L., and H.O.S. conceptualized the study. H.Y.L., G.A., Y.J., S.M., J.L., W.-S.H., W.H., and H.O.S. developed the methodology. G.A., Y.J., S.M., W.H., J.C., S.L., J.L., W.-S.H., S.K., Y.K., and H.O.S. developed the software. H.Y.L., G.A., Y.J., E.J.S., S.M., J.L., W.-S.H., W.H., and H.O.S. performed the validation. H.Y.L., G.A., S.M., J.L., W.-S.H., W.H., and H.O.S. conducted the formal analyses. H.Y.L., G.A., Y.J., E.J.S., H.G.R., S.M., H.-L.Y., H.-C.L., J.C., S.L., J.L., W.-S.H., W.H., and H.O.S. contributed to the investigation and data curation. H.Y.L., H.G.R., B.W.P., V.H., D.A.D., J.C., S.L., S.-M.L., and H.O.S. provided resources. H.Y.L., G.A., W.H., and H.O.S. prepared the visualizations. H.Y.L., G.A., Y.J., E.J.S., J.C., R.R., and H.O.S. drafted the original manuscript. All authors reviewed and edited the manuscript. H.Y.L., S.-M.L., and H.O.S. acquired funding. H.Y.L., H.G.R., J.C., S.-M.L., and H.O.S. administered the project. E.J.S., B.W.P., V.H., D.A.D., R.R., K.C., S.-M.L., and H.O.S. supervised the study. All authors read and approved the final manuscript.
Funding
This research was supported by a grant from the Korea Health Technology R&D Project through the Korea Health Industry Development Institute, funded by the Ministry of Health and Welfare, Republic of Korea (grant number RS-2021-KH114277).
The funder had no role in the design or conduct of the study; collection, management, analysis, or interpretation of the data; preparation, review, or approval of the manuscript; or the decision to submit the manuscript for publication.
Data availability
The SNUH data analyzed during the current study are available from the corresponding author on reasonable request. The Mayo Clinic data were accessed in de-identified form through the Mayo Clinic Platform under applicable data-governance and privacy restrictions and cannot be redistributed by the authors.
Declarations
Ethics approval and consent to participate
This study was conducted in accordance with the Declaration of Helsinki and the principles of Good Clinical Practice. The Institutional Review Board of Seoul National University Hospital approved the study and the use of the Seoul National University Hospital data (approval number H-2109-143-1258). Because this was a retrospective observational analysis of de-identified data, the requirement for written informed consent was waived by the Institutional Review Board.
The Mayo Clinic data were accessed in a de-identified form through the Mayo Clinic Platform under applicable institutional data-governance and privacy-preserving procedures. The data were determined to be de-identified in accordance with the Health Insurance Portability and Accountability Act Privacy Rule.
Consent for publication
Not applicable. The manuscript does not contain identifiable individual patient information.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Hong Yeul Lee and Gaon An have equally contributed as first authors.
Sang-Min Lee and Hyun Oh Song have equally contributed as corresponding authors.
Contributor Information
Sang-Min Lee, Email: sangmin2@snu.ac.kr.
Hyun Oh Song, Email: hyunoh@snu.ac.kr.
References
- 1.Urner M, Jüni P, Hansen B, Wettstein MS, Ferguson ND, Fan E. Time-varying intensity of mechanical ventilation and mortality in patients with acute respiratory failure: a registry-based, prospective cohort study. Lancet Respir Med. 2020;8:905–13. 10.1016/S2213-2600(20)30325-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Peñuelas O, Keough E, López-Rodríguez L, Carriedo D, Gonçalves G, Barreiro E, et al. Ventilator-induced diaphragm dysfunction: translational mechanisms lead to therapeutical alternatives in the critically ill. Intensive Care Med Exp. 2019;7:48. 10.1186/s40635-019-0259-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Seo Y, Lee H-J, Ha EJ, Ha TS. 2021 KSCCM clinical practice guidelines for pain, agitation, delirium, immobility, and sleep disturbance in the intensive care unit. Acute Crit Care. 2022;37:1–25. 10.4266/acc.2022.00094. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Goligher EC, Dres M, Patel BK, Sahetya SK, Beitler JR, Telias I, et al. Lung- and diaphragm-protective ventilation. Am J Respir Crit Care Med. 2020;202:950–61. 10.1164/rccm.202003-0655CP. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Qadir N, Sahetya S, Munshi L, Summers C, Abrams D, Beitler J, et al. An update on management of adult patients with acute respiratory distress syndrome: an official American Thoracic Society clinical practice guideline. Am J Respir Crit Care Med. 2024;209:24–36. 10.1164/rccm.202311-2011ST. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Pelosi P, Ball L, Barbas CSV, Bellomo R, Burns KEA, Einav S, et al. Personalized mechanical ventilation in acute respiratory distress syndrome. Crit Care Lond Engl. 2021;25:250. 10.1186/s13054-021-03686-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Nguyen Y-L, Perrodeau E, Guidet B, Trinquart L, Richard J-CM, Mercat A, et al. Mechanical ventilation and clinical practice heterogeneity in intensive care units: a multicenter case-vignette study. Ann Intensive Care. 2014;4:2. 10.1186/2110-5820-4-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Colombo D, Cammarota G, Bergamaschi V, De Lucia M, Corte FD, Navalesi P. Physiologic response to varying levels of pressure support and neurally adjusted ventilatory assist in patients with acute respiratory failure. Intensive Care Med. 2008;34:2010–8. 10.1007/s00134-008-1208-3. [DOI] [PubMed] [Google Scholar]
- 9.Akbulut FP, Akkur E, Akan A, Yarman BS. A decision support system to determine optimal ventilator settings. BMC Med Inform Decis Mak. 2014;14:3. 10.1186/1472-6947-14-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Sinderby C, Navalesi P, Beck J, Skrobik Y, Comtois N, Friberg S, et al. Neural control of mechanical ventilation in respiratory failure. Nat Med. 1999;5:1433–6. 10.1038/71012. [DOI] [PubMed] [Google Scholar]
- 11.Younes M. Proportional assist ventilation, a new approach to ventilatory support. Theory. Am Rev Respir Dis. 1992;145:114–20. 10.1164/ajrccm/145.1.114. [DOI] [PubMed] [Google Scholar]
- 12.Ceylan G, Topal S, Atakul G, Colak M, Soydan E, Sandal O, et al. Randomized crossover trial to compare driving pressures in a closed-loop and a conventional mechanical ventilation mode in pediatric patients. Pediatr Pulmonol. 2021;56:3035–43. 10.1002/ppul.25561. [DOI] [PubMed] [Google Scholar]
- 13.Dojat M, Pachet F, Guessoum Z, Touchard D, Harf A, Brochard L. NéoGanesh: a working system for the automated control of assisted ventilation in ICUs. Artif Intell Med. 1997;11:97–117. 10.1016/s0933-3657(97)00025-0. [DOI] [PubMed] [Google Scholar]
- 14.Otis AB, Fenn WO, Rahn H. Mechanics of breathing in man. J Appl Physiol. 1950;2:592–607. 10.1152/jappl.1950.2.11.592. [DOI] [PubMed] [Google Scholar]
- 15.Mead J. Control of respiratory frequency. J Appl Physiol. 1960;15:325–36. 10.1152/jappl.1960.15.3.325. [DOI] [Google Scholar]
- 16.Sinnige JS, Buiteman-Kruizinga LA, Horn J, Paulus F, Schultz MJ, Serpa Neto A, et al. Effect of automated closed-loop ventilation vs protocolized conventional ventilation on ventilator-free days in critically Ill adults: a randomized clinical trial. JAMA. 2026;335:874–84. 10.1001/jama.2025.24384. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Ouyang L, Wu J, Jiang X, Almeida D, Wainwright C, Mishkin P, et al. Training language models to follow instructions with human feedback. Adv Neural Inf Process Syst [Internet]. 2022 [cited 2024 May 2]. p. 27730–44. https://proceedings.neurips.cc/paper_files/paper/2022/hash/b1efde53be364a73914f58805a001731-Abstract-Conference.html. Accessed 2 May 2024
- 18.An G, Lee J, Zuo X, Kosaka N, Kim K-M, Song HO. Direct Preference-based Policy Optimization without Reward Modeling. Adv Neural Inf Process Syst [Internet]. 2023 [cited 2024 May 2]. p. 70247–66. https://proceedings.neurips.cc/paper_files/paper/2023/hash/de8bd6b2b01cfa788e63f62e5b9a99b9-Abstract-Conference.html. Accessed 2 May 2024
- 19.Lee H-C, Jung C-W. Vital recorder—a free research tool for automatic recording of high-resolution time-synchronised physiological data from multiple anaesthesia devices. Sci Rep. 2018;8:1527. 10.1038/s41598-018-20062-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Kostrikov I, Nair A, Levine S. Offline Reinforcement Learning with Implicit Q-Learning [Internet]. arXiv; 2021 [cited 2025 Dec 31]. 10.48550/arXiv.2110.06169 [DOI]
- 21.Snell C, Kostrikov I, Su Y, Yang M, Levine S. Offline RL for Natural Language Generation with Implicit Language Q Learning [Internet]. arXiv; 2023 [cited 2025 Dec 31]. 10.48550/arXiv.2206.11871 [DOI]
- 22.Voloshin C, Le HM, Jiang N, Yue Y. Empirical study of off-policy policy evaluation for reinforcement learning [Internet]. arXiv; 2021 [cited 2025 Dec 31]. 10.48550/arXiv.1911.06854 [DOI]
- 23.Yehya N, Harhay MO, Curley MAQ, Schoenfeld DA, Reeder RW. Reappraisal of ventilator-free days in critical care research. Am J Respir Crit Care Med. 2019;200:828. 10.1164/rccm.201810-2050CP. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Kirakli C, Naz I, Ediboglu O, Tatar D, Budak A, Tellioglu E. A randomized controlled trial comparing the ventilation duration between adaptive support ventilation and pressure assist/control ventilation in medical patients in the ICU. Chest. 2015;147:1503–9. 10.1378/chest.14-2599. [DOI] [PubMed] [Google Scholar]
- 25.Peine A, Hallawa A, Bickenbach J, Dartmann G, Fazlic LB, Schmeink A, et al. Development and validation of a reinforcement learning algorithm to dynamically optimize mechanical ventilation in critical care. NPJ Digit Med Nature Publishing Group. 2021;4:1–12. 10.1038/s41746-021-00388-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Kondrup F, Jiralerspong T, Lau E, Lara N de, Shkrob J, Tran MD, et al. Towards safe mechanical ventilation treatment using deep offline reinforcement learning. Proc AAAI Conf Artif Intell [Internet]. 2023 [cited 2024 May 3]. p. 15696–702. 10.1609/aaai.v37i13.26862 [DOI]
- 27.Prasad N, Cheng L-F, Chivers C, Draugelis M, Engelhardt BE. A Reinforcement Learning Approach to Weaning of Mechanical Ventilation in Intensive Care Units [Internet]. arXiv; 2017 [cited 2024 May 3]. 10.48550/arXiv.1704.06300 [DOI]
- 28.Johnson AEW, Pollard TJ, Shen L, Lehman LH, Feng M, Ghassemi M, et al. MIMIC-III, a freely accessible critical care database. Sci Data. 2016;3:160035. 10.1038/sdata.2016.35. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10:1. 10.1038/s41597-022-01899-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Pollard TJ, Johnson AEW, Raffa JD, Celi LA, Mark RG, Badawi O. The eICU collaborative research database, a freely available multi-center database for critical care research. Sci Data. 2018;5:180178. 10.1038/sdata.2018.178. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Faltys M, Zimmermann M, Lyu X, Hüser M, Hyland S, Rätsch G, et al. HiRID, a high time-resolution ICU dataset [Internet]. PhysioNet; [cited 2026 July 25]. 10.13026/NKWC-JS72 [DOI]
- 32.Christiano PF, Leike J, Brown T, Martic M, Legg S, Amodei D. Deep reinforcement learning from human preferences. Adv Neural Inf Process Syst [Internet]. Curran Associates, Inc.; 2017 [cited 2024 May 2]. https://proceedings.neurips.cc/paper_files/paper/2017/hash/d5e2c0adad503c91f91df240d0cd4e49-Abstract.html. Accessed 2 May 2024
- 33.Skalse J, Howe N, Krasheninnikov D, Krueger D. Defining and characterizing reward gaming. Adv Neural Inf Process Syst [Internet]. 2022 [cited 2024 May 2]. p. 9460–71. https://proceedings.neurips.cc/paper_files/paper/2022/hash/3d719fee332caa23d5038b8a90e81796-Abstract-Conference.html. Accessed 2 May 2024
- 34.Goossen RL, Schultz MJ, Tschernko E, Chew MS, Robba C, Paulus F, et al. Effects of closed loop ventilation on ventilator settings, patient outcomes and ICU staff workloads—a systematic review. Eur J Anaesthesiol. 2024;41:438–46. 10.1097/EJA.0000000000001972. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Rose L, Schultz MJ, Cardwell CR, Paulus F, Couper K, Jouvet P, et al. Automated versus non-automated weaning for reducing the duration of mechanical ventilation for critically ill adults and children. Cochrane Database Syst Rev. 2025;7:CD009235. 10.1002/14651858.CD009235.pub4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Zhang B, Ratano D, Brochard LJ, Georgopoulos D, Duffin J, Long M, et al. A physiology-based mathematical model for the selection of appropriate ventilator controls for lung and diaphragm protection. J Clin Monit Comput. 2021;35:363–78. 10.1007/s10877-020-00479-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Semler MW, Casey JD, Lloyd BD, Hastings PG, Hays MA, Stollings JL, et al. Oxygen-saturation targets for critically ill adults receiving mechanical ventilation. N Engl J Med. 2022;387:1759–69. 10.1056/NEJMoa2208415. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Chu DK, Kim LH-Y, Young PJ, Zamiri N, Almenawer SA, Jaeschke R, et al. Mortality and morbidity in acutely ill adults treated with liberal versus conservative oxygen therapy (IOTA): a systematic review and meta-analysis. Lancet. 2018;391:1693–705. 10.1016/S0140-6736(18)30479-3. [DOI] [PubMed] [Google Scholar]
- 39.Fan E, Del Sorbo L, Goligher EC, Hodgson CL, Munshi L, Walkey AJ, et al. An official American Thoracic Society/European Society of Intensive Care Medicine/Society of Critical Care Medicine clinical practice guideline: mechanical ventilation in adult patients with acute respiratory distress syndrome. Am J Respir Crit Care Med. 2017;195:1253–63. 10.1164/rccm.201703-0548ST. [DOI] [PubMed] [Google Scholar]
- 40.Brower RG, Lanken PN, MacIntyre N, Matthay MA, Morris A, Ancukiewicz M, et al. Higher versus lower positive end-expiratory pressures in patients with the acute respiratory distress syndrome. N Engl J Med. 2004;351:327–36. 10.1056/NEJMoa032193. [DOI] [PubMed] [Google Scholar]
- 41.Pletsch-Assuncao R, Caleffi Pereira M, Ferreira JG, Cardenas LZ, de Albuquerque ALP, de Carvalho CRR, et al. Accuracy of invasive and noninvasive parameters for diagnosing ventilatory overassistance during pressure support ventilation. Crit Care Med. 2018;46:411–7. 10.1097/CCM.0000000000002871. [DOI] [PubMed] [Google Scholar]
- 42.Rebello A, Tang S, Wiens J, Parbhoo S. Leveraging factored action spaces for off-policy evaluation [Internet]. arXiv; 2023 [cited 2026 Aug 28]. 10.48550/arXiv.2307.07014 [DOI]
- 43.Tang S, Makar M, Sjoding MW, Doshi-Velez F, Wiens J. Leveraging factored action spaces for efficient offline reinforcement learning in healthcare [Internet]. arXiv; 2023 [cited 2026 Aug 28]. 10.48550/arXiv.2305.01738 [DOI]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The SNUH data analyzed during the current study are available from the corresponding author on reasonable request. The Mayo Clinic data were accessed in de-identified form through the Mayo Clinic Platform under applicable data-governance and privacy restrictions and cannot be redistributed by the authors.