Introduction

Effectively retrieving patient information is vital for patient care and safety in the intensive care unit (ICU), in which multiple information systems and devices are used for monitoring and treatment. Critically ill patients undergo various tests and procedures daily, with a limited physiological reserve to tolerate errors1. Inadequate retrieval and documentation of patient information as well as restricted access to this data have been identified as significant barriers to effective information exchange in the ICU2,3 leading to errors in patient care decisions and subsequent patient harm4. Thus, efficient and accurate data management is crucial for informed decision-making during patient rounds in the ICU5.

However, data retrieval and organization by ICU clinicians are often incomplete, including frequent omissions and inaccuracies that go largely unnoticed by the primary care team, despite the use of a well-established electronic medical record (EMR) system4. This can be attributed to the complexity of the ICU environment, where the vast and continually growing volume of information—especially in patients receiving ICU interventions such as mechanical ventilation, inotrope or vasoactive medications, extracorporeal membrane oxygenation (ECMO), or hemodialysis—further complicates data collection and organization. The volume of documented information may correlate with therapeutic intensity, and such extensive documentation increases workload and opportunities for error6,7. Information overload has become a significant issue in the ICU, as critical care providers are often overwhelmed by the sheer volume of data. Organizing such information in the ICU, particularly dynamic quantitative data, can potentially lead to cognitive fatigue8,9. Physicians in the ICU utilize only a limited set of clinical information, whereas the EMR contains a vast amount of unused data10.

To address these issues, ICU patient-centric Internet of Things (IoT)-enabled systems in healthcare have been proposed to help transmit large volumes of real-time data from numerous medical devices, enable continuous and efficient monitoring, enhance the accuracy of data collection and analysis11,12and potentially minimize errors13. Recent advancements in machine learning and artificial intelligence (AI), coupled with IoT infrastructure, indicate that artificial intelligence of things (AIoT) technology will be pivotal in driving healthcare innovations14,15.

This pilot study aims to compare the efficiency and accuracy of an AIoT-enabled ICU command center (CC)16 with traditional data collection methods utilizing the existing hospital information system (HIS) and seeks to reveal how IoT technologies can enhance data management for healthcare provider (HCP) in the ICU.

Method

Ethical approval

The study was approved by the Institutional Review Board of CMUH (CMUH111-REC2-220) and was conducted in accordance with the principles of the Declaration of Helsinki. Informed consent was obtained from all participants.

Study design

This study was a randomized, two-sequence crossover pilot study conducted on April 19, 2024, in a 20-bed respiratory intensive care unit (RICU) at China Medical University Hospital (CMUH), a tertiary referral hospital in Taiwan. Each participating HCP collected data from the same five ICU patients using both the traditional HIS and the ICU CC systems, following either the CC-HIS or HIS-CC sequence (Fig. 1). The sequence assignment was randomized using a lottery method. This ensured that each HCP used both methods on the same patients, thereby controlling for individual variability between the HIS and CC groups.

Fig. 1
figure 1

Consort flowchart of the participants. aOut of six patients, HCPs gathered data from five and subsequently collected data from the same five patients using either the CC-HIS or HIS-CC sequence. The order of the patients was randomized. bAfter excluding 26 sets of omitted or insufficient data, 184 datasets were available for analysis, with 92 sets using CC and 92 sets using HIS. Abbreviations: CC command center, HIS hospital information system, ICU intensive care unit.

Full size image

Participants and eligibility criteria

Eligible participants were HCPs with ICU training experience, including nurse practitioners (NPs), residents, fellows, and attending physicians. Prior to the study, all providers received a standardized instructions and training on the AIoT-enable CC platform. To reflect the typical daily workload in the study RICU, each HCP collected data from five patients. Participants were excluded from analysis if they declined participation, withdrew before completing the intervention, or had data recorded using an incorrect method.

Interventions

Data collection is believed to be more challenging in critically ill patients6; therefore, the study enrolled the six most critical ICU patients on the study date. These patients were selected based on their Acute Physiologic Assessment and Chronic Health Evaluation-II (APACHE II) scores. All HCPs performed the same standardized data collection task using a structured daily ICU checklist (Supplementary Table 1), regardless of their clinical background or role. This checklist is designed for comprehensive prerounding data gathering and is routinely used in the CMUH RICU. The use of a fixed patient set, and standardized task design helped isolate the impact of the data platform (AIoT CC vs. HIS), minimizing confounding effects related to workflow complexity or clinician experience. The order of the patients was randomized using a computer-generated random sequence to minimize bias (Supplementary Table 2). There was no washout period between the systems because the study also aimed to observe how collecting more data affected time requirements and data accuracy.

Hospital information system

The HIS at CMUH adopts a web-based user environment that utilizes Health Level Seven standards and Extensible Markup Language to manage the complex medical environment, like most HISs in Taiwan17,18. The HIS consists of various heterogeneous systems, including computerized physician order entry, EMR, laboratory information systems, nursing information systems, nutrition information systems, pharmacy information systems, respiratory care information systems, and the picture archiving and communication system.

Figure 2 illustrates the sitemap for ICU data retrieval from the HIS of CMUH. Although the HIS provides comprehensive and detailed information and is designed to facilitate integration across various systems, the ICU environment presents unique challenges. Most devices, including patient monitors, mechanical ventilators, drug infusion pumps, and hemodynamic monitors, are unable to transmit data effectively. Moreover, to gain a comprehensive understanding of a patient’s condition, HCP must navigate through multilayered systems to collect data from disparate sources.

Fig. 2
figure 2

Data retrieval sitemap of the HIS at CMUH. HCP must navigate through multilayered systems to collect data from disparate sources. aDisease severity was assessed using the Acute Physiologic Assessment and Chronic Health Evaluation-II (APACHE II). bPain score was assessed using CCPOT, Faces Pain Scale, and NRS. cAgitation–sedation status was assessed using RASS. dDelirium was assessed using ICDSC. eCatheters, including arterial lines, central venous catheters, double-lumen catheters, nasogastric tubes, Foley catheters, and their usage days. fMedical devices whose data could not be transmitted directly into the HIS were recorded by the nursing staff, including drug infusion pumps, hemodynamic monitors, CRRT, and ECMO. gNutrition data that could not be transmitted directly into the HIS were recorded by the nutritionist. hMechanical ventilator data that could not be transmitted directly into the HIS were recorded by the respiratory therapist. Abbreviations: CPOE computerized physician order entry, EMR electronic medical records, HIS hospital information systems, ICDSC Intensive Care Delirium Screening Checklist, I/O intake and output, LIS laboratory information systems, MV mechanical ventilator, Nursing IS nursing information systems, Nutrition IS nutrition information systems, PPN peripheral parental nutrition, RASS Richmond Agitation–Sedation Scale, RCIS respiratory care information systems, PACS picture archiving and communication system, TPN total parental nutrition.

Full size image

Construction and design of an AIoT-enabled ICU command center

The AIoT-enabled ICU CC at CMUH is built on a four-layer AIoT platform that adopts the international Fast Healthcare Interoperability Resources standard as its data format to facilitate seamless electronic data exchange across disparate healthcare systems16. The AIoT platform integrates AI with medical IoT devices to enable real-time, scalable data integration, processing, and intelligent presentation. Figure 3 A illustrates the architecture of the AIoT-enabled ICU CC. It encompasses three main components: (1) data acquisition, which collects information from medical devices, the HIS, AI and business intelligence (BI) servers, and real-time cameras; (2) data transformation, which involves sanitizing and standardizing data formats to comply with international standards; and (3) data presentation, which organizes the processed data into structured dashboards for visualization and AI- and BI-supported decision-making. The initial version of the AIoT-enabled CC at CMUH includes four AI models designed for the early detection and preemptive management of acute respiratory distress syndrome (ARDS)19 sepsis20 drug-resistant pathogens21,22,23 and ST-elevation myocardial infarction24. Additionally, it features a BI tool that provides data-driven support for monitoring lung-protective strategies in patients with ARDS25.

Figure 3B shows the web-based user interface of the personal panel in the ICU CC. This centralized dashboard consolidates key patient information, reducing the need to navigate multiple systems. Integrated 3D body diagrams and graphics help clinicians quickly interpret complex data and identify trends or abnormalities. Unlike conventional HIS platforms, which rely on static data retrieval and fragmented interfaces, the AIoT-based system integrates data from multiple sources, ensuring data integrity and delivering real-time, continuous, and dynamic information.

Fig. 3
figure 3

(A) System architecture of the AIoT-enabled ICU CC. (1) FHIR Translator and (2) Load Balancer: Convert and standardize data from medical devices and EMRs into the FHIR format, then distribute it across parallel data pipelines. (3) Message Queue (MQ): A distributed system that shards and stores real-time data (e.g., from patient monitors) across multiple nodes. (4) Historical Database: Stores cold data for long-term tracking and analysis. (5) WebRTC Server: Streams real-time video from ICU cameras. (6) Consumer Dispatcher: Delivers curated, event-driven data to Tele-ICU, central stations, and AIoT applications. Adapted from our previous work16 (B) Personal panel of the ICU CC. Centralizes critical patient information into a single dashboard, providing a comprehensive overview and reducing the need to switch between multiple systems or documents. User interface designed by our research team using Figma (web-based application, accessed on 11 August 2025, https://www.figma.com/). Abbreviations: AI artificial intelligence, BI business intelligence, EMR electronic medical records, FHIR fast healthcare interoperability resources, HIS hospital information system, ICU intensive care unit, LIS laboratory information systems, PACS picture archiving and communication system, PM patient monitor, VEN Ventilator, WebRTC web real-time communication.

Full size image

Measurements

Data collection was performed using a standardized and structured ICU checklist designed to support goal-oriented understanding of patient care (Supplementary Table 1). Data collection time and accuracy were recorded for each patient on both platforms. Accuracy was defined as the percentage of correct information, with incorrect or omitted data assessed by comparing the information recorded in the checklist with the actual patient conditions in the EMR. The percentage Change in the data collection time was defined as 100% × (CC – HIS)/HIS. Participant characteristics (the professional titles, ICU full-time status, years of experience) and patient characteristics (APACHE II score, ECMO, CRRT, number of infusion pumps and catheters) were documented. ICU full-time status was defined as healthcare professionals who primarily work in the ICU, while non-full-time staff are those who are not employed full-time in the ICU but may rotate through the ICU for training or on-duty assignments.

Sample size estimation

As a pilot study, there were no prior trials comparing the AIoT and HIS systems in terms of data collection efficiency and accuracy. However, preliminary testing before the trial revealed a difference of approximately 3 min in data collection time and a 1% difference in data accuracy between the CC and HIS. Based on these findings, a sample size estimation using a two-sided t-test with a 5% type I error rate indicated that 46 pairs would be necessary to achieve a power of 90% to detect a mean difference of 3 min in time and 1% in accuracy, assuming standard deviations of 5 min for time and 2% for accuracy, respectively. Additionally, the study explored how collecting more data affected time requirements and data accuracy. A crossover design, in which eligible HCP collected data from the same five patients, was implemented to ensure the required sample size.

Outcome variables

The primary outcomes were efficiency (measured by data collection time) and accuracy (measured by the percentage of correct data). The secondary outcomes included a further analysis of efficiency and accuracy across different subgroups as well as the impact of increased data collection (measured by the order of data collection) on time requirements and data accuracy.

Statistical analysis

All statistical analyses were performed using SAS version 9.4 (SAS Institute, Cary, NC, USA). Continuous variables are presented as the mean and standard deviations (SD). A generalized linear model (GLM) with generalized estimating equation (GEE) was used to adjust for the crossover design and repeated measures within participants and to estimate the differences between the CC and HIS. In the GLM, the strength of the associations was expressed using parameter estimates with 95% confidence intervals. All tests were two-sided, and a p-value of < 0.05 was considered statistically significant.

Results

Participant characteristics

A total of 23 HCPs were randomized into the study. Thirteen HCPs were allocated to the CC-HIS group. However, one HCP did not receive the allocated intervention due to a schedule change, and another HCP was excluded from the analysis due to incorrect data recording methods. Ten HCPs were allocated to the HIS-CC group (Fig. 1).

From six ICU patients, HCP gathered data from the same five patients using either the CC-HIS or HIS-CC sequence. The demographic data of the enrolled providers are presented in Table 1. NPs and residents comprised 67.3% of the cohort. Additionally, 72.8% of the HCPs had less than six years of work experience and 59.1% were not full-time ICU staff.

Table 1 Demographic data of the enrolled healthcare provider.
Full size table

The clinical characteristics of the ICU patients from whom data were gathered are summarized in Table 2. The patients aged 57–84 years were primarily admitted to the ICU because of ARDS, septic shock, or out-of-hospital cardiac arrest. The severity of their conditions, reflected by APACHE II scores, ranged from 35 to 50 on the day of data collection. Two patients received ECMO, whereas another required CRRT. Additionally, the number of infusion pumps used varied from one to four, and the number of catheters ranged between four and six, highlighting the complexity of the clinical scenario. A total of 210 patient data sets were initially recorded by the HCPs. After excluding those with missing or insufficient data, 184 sets remained for analysis, evenly divided between 92 sets utilizing CC and 92 sets utilizing HIS.

Table 2 Clinical characteristics of the enrolled patients in the ICU.
Full size table

Efficiency and accuracy

The CC system significantly reduced data collection time by 41.8% compared with the HIS (p < 0.0001), with an average reduction of 6.75 min per patient (8.5 vs. 15.2 min, p < 0.0001) (Table 3). This reduction was particularly notable in cases involving more infusion pumps, higher APACHE II scores, and among more experienced and full-time ICU staff. Additionally, the CC system significantly improved data accuracy by 2.07% compared with the HIS group (95.9% vs. 93.8%, p = 0.0002) (Table 3). The improvement in accuracy was more pronounced among nonattending HCP and non-ICU full-time staff. Figure 4 demonstrates that the CC system consistently achieved higher accuracy with shorter data collection times, along with less variability in both accuracy and time, reflecting more efficient and consistent performance compared to the HIS group.

Table 3 Comparison of efficiency and accuracy in data collection using the ICU CC and HIS.
Full size table
Fig. 4
figure 4

Correlation between efficiency and accuracy for the ICU CC and HIS. The scatter plot illustrates the relationship between accuracy (x-axis) and data collection time (y-axis) for the two groups. The CC group (orange dots) tended to show higher accuracy with shorter data collection times, indicating more efficient performance. In contrast, the HIS group (blue dots) exhibited a wider spread in both accuracy and data collection times, with some data points showing lower accuracy and longer data collection times. This suggests that the CC group is generally more efficient and consistent in data collection than the HIS group. Abbreviations: CC command center, ICU intensive care unit, HIS hospital information system.

Full size image

Clinical factors predicting the efficiency and accuracy of data collection

In the GLM with GEE analysis, the use of the CC system was significantly associated with decreased data collection time (estimate: − 6.75 min, 95% CI: − 7.68 to − 5.81, p < 0.0001) and improved data accuracy (estimate: +2.07%, 95% CI: 0.99 to 3.14, p = 0.0002), compared to the HIS. An APACHE II score greater than 38 was correlated with increased data collection time (estimate: +1.41 min, 95% CI: 0.03 to 2.78, p = 0.0450) but was not significantly associated with changes in data accuracy (p = 0.8829).

Regarding the percent change in data collection time, attending physicians (estimate: − 13.48%, 95% CI: − 24.09 to − 2.87, p = 0.0128) and ICU full-time staff (estimate: − 13.68%, 95% CI: − 23.88 to − 3.49, p = 0.0085) demonstrated significant reductions when using the CC system.

In terms of data accuracy, significant improvements were associated with the use of the CC system (estimate: +2.07%, 95% CI: 0.99 to 3.14, p = 0.0002), being an attending physician (estimate: +5.19%, 95% CI: 2.17 to 8.20, p = 0.0008), and being ICU full-time staff (estimate: +3.02%, 95% CI: 0.24 to 5.80, p = 0.0334). Conversely, having more than 6 years of work experience was associated with reduced data accuracy (estimate: − 7.49%, 95% CI: − 11.03 to − 3.96, p < 0.0001) (Table 4).

Table 4 Generalized linear model comparing the efficiency and accuracy of data collection across different variables.
Full size table

The effect of increased data collection on time requirements and data accuracy

The trend line of the mean data collection time shows a decline with the sequential order of data collection in both groups. The HIS demonstrated a more pronounced reduction in the data collection time, whereas the CC exhibited a more gradual decrease (Fig. 5A). Both groups showed a similar decline in accuracy with the sequential order of data collection; however, the CC consistently maintained a higher level of accuracy than the HIS throughout the data collection sequence (Fig. 5B).

Fig. 5
figure 5

Learning curves and fatigue effects between the ICU CC and HIS (A) Learning curves associated with increased data collection for the ICU CC and HIS. The linear regression lines for the CC and HIS illustrate the relationship between the sequence of data collection (x-axis) and the mean data collection time (y-axis). Overall, the trend line of the mean data collection time decreased with the sequential order of data collection in both the CC and HIS groups. The HIS group exhibited a more pronounced learning effect, as indicated by a steeper decline in data collection time, suggesting greater improvement with an increased data collection workload. In contrast, the CC group showed a more gradual decline, implying that the CC system is more user-friendly and requires less effort to master. (B) Fatigue effect on data accuracy associated with increased data collection for the ICU CC and HIS. The linear regression lines for the CC and HIS illustrate the relationship between the sequence of data collection (x-axis) and data accuracy (y-axis). Both groups exhibited a similar decline in accuracy with the sequential order of data collection, indicating a fatigue effect where performance decreased with repetitive data collection. Despite this, the CC consistently maintained a higher level of accuracy compared to the HIS throughout the data collection sequence. Abbreviations: CC command center, ICU intensive care unit, HIS hospital information system.

Full size image

Discussion

This pilot study demonstrates that even in patients with high disease severity, the AIoT-enabled ICU CC system reduced data collection time while improving data accuracy compared with the existing, well-established HIS. The reduction in data collection time was more pronounced among full-time ICU staff, but improvements in data accuracy were more notable among non-ICU full-time staff. These findings provide insights into how IoT technologies can enhance data management for HCP in the ICU and highlight the potential for further improving patient outcomes in critical care.

This study revealed a significant 41.8% reduction in data collection time when utilizing the CC system, with the effect being more pronounced among attending physicians and full-time ICU staff. This may be attributed to their greater experience in quickly identifying relevant information, allowing them to benefit more from the structured, integrated display offered by the CC system. This finding underscores the value of the CC system in supporting healthcare professionals who must frequently make timely decisions in a high-stakes environment. ICU staff shortages and work overload have presented unprecedented challenges to the healthcare system, a problem that has been especially significant during the SARS-CoV-2 pandemic26. Thus, finding strategies to boost efficiency, alleviate workload, and prevent burnout among ICU staff is essential. Nadkarni et al. emphasized the importance of a user-friendly system for healthcare professionals and the standardization of messaging protocols across different platforms to enhance the utilization of data in critical care14. Previous studies have shown that an organized, visualized, and patient-centered dashboard is associated with improved efficiency in clinical data management in the ICU8,27. However, these studies relied on conventional data integration processes from existing EMR systems. In contrast, we developed an IoT platform that directly transmits extensive real-time data streams from multiple medical IoT devices, enhancing interoperability and allowing clinicians to manage substantial amounts of medical data with minimal delay and at accelerated speeds.

We found that a higher APACHE II score were correlated with increased data collection time among the relatively high-severity, critically ill patients in our cohort. Additionally, these patients required extensive critical care resources, including the use of CRRT, ECMO, and up to four infusion pumps. Previous studies have highlighted that the volume of information documented during critical illness increases significantly, especially for patients requiring ventilators, inotrope or vasoactive medications, hemodialysis, or ECMO6leading ICU physicians to spend more time using EMR to gather data on these patients7. Despite the complexity of the critical patients, the CC system substantially reduced data collection time in patients with more infusion pumps and higher APACHE II scores. This suggests that the system is particularly beneficial in managing more complex disease progressions.

The use of the CC system significantly improved data accuracy, particularly among nonattending and non-ICU full-time staff. Attending physicians and full-time ICU staff are likely more familiar with retrieving data from the HIS across different systems, which may explain why the improvement in data accuracy was less pronounced for them. However, the primary care workforce in this study was predominantly composed of nonattending and non-ICU full-time staff. This finding aligns with previous studies showing that although physicians with more clinical experience integrate data more successfully, the majority of data collection tasks often fall to the least experienced members of the team4,28. Data accuracy is crucial for making informed decisions in the ICU, where even small errors can have severe consequences1,14. Our findings underscore the effectiveness of the CC system in improving performance and reducing errors among HCP who are not full-time ICU staff but rotate through the ICU for training or on-duty assignments, which may further enhance patient care.

In this study, the CC system exhibited a smaller learning curve than the HIS, as shown by a less pronounced decline in data collection time under increased workload. This suggests that the CC system is more user-friendly and requires less effort to master. Although both groups showed a similar decline in accuracy with increased workload, the CC group consistently maintained a higher level of accuracy than the HIS group throughout the data collection process. Information overload is a widespread and serious problem in the ICU6where clinicians manage myriad pieces of ICU information and suffer from cognitive fatigue8,9. Notably, visual presentation of patient data has been shown to improve data display and reduce cognitive overload among clinicians3,8,9,29. Moreover, advancements in machine learning algorithms and their integration with clinical decision support systems hold promise for reducing the workload of ICU clinicians and enhancing patient care14. From this perspective, the AIoT-enabled CC system, which integrates the current EMR system, AI models, and a BI platform, may help mitigate cognitive fatigue, sustain higher HCP performance, and exemplify the potential of future AIoT-driven intelligent ICU systems.

This study has limitations that warrant consideration. First, although the sample size is sufficient for a pilot study, it may not capture the full variability of the ICU settings. In addition, when patients were categorized for subgroup analysis (e.g., number of pumps, ECMO use, or CRRT use), some categories contained relatively small case numbers, resulting in substantial data imbalance and reduced statistical power in subgroup comparisons. Second, the checklist captured only limited information on CRRT and ECMO. Specifically, only the ECMO type and whether CRRT was used were recorded, without detailed device settings. This may have reduced the accuracy of reflecting illness severity in these subgroup analysis. Third, since we selected the most critically ill patients, the difference in performance between the CC and HIS systems among less severely ill ICU patients remains unclear. Additionally, the crossover design, while robust in controlling for individual differences, may introduce a learning effect that could influence the results. Future studies with larger and more diverse populations, as well as long-term evaluations are necessary to fully understand the impact of the CC system on ICU operations and patient outcomes. Finally, the effects of reducing the time spent on cognitive processing by HCP are unknown, as previous study have suggested that decreasing the time spent on data collection have unintended negative effects on clinicians’ cognitive processing4. The future application of this technology in clinical practice should be carefully evaluated.

Conclusions

In conclusion, compared to the existing HIS, the AIoT-enabled ICU CC significantly enhances efficiency by reducing data collection time and improving data accuracy, suggesting its potential to address the challenges of information overload in the ICU. Future studies are warranted to evaluate the impact of these findings on patient outcomes.