Real-World Deployment and Operational Evaluation of an On-Premise Thermal IoMT System in an Intensive Care Unit

Abstract

Thermal imaging provides a non-contact approach for continuous environmental monitoring in intensive care units, although most existing studies remain limited to controlled or experimental settings with limited real-world validation. This study aimed to design, deploy, and operationally evaluate a scalable, low-cost, on-premise thermal Internet of Medical Things system in a real intensive care unit. A distributed network of seven thermal acquisition nodes, scalable to thirteen beds, was implemented using embedded devices and radiometric infrared cameras. Thermal images were transmitted through the hospital Wi-Fi network to a local server for storage, indexing, and near real-time visualization. System performance was evaluated using operational metrics including end-to-end latency, throughput, availability, and connectivity incidents. During a 10-day structured evaluation, the system generated 5,763,312 thermal images (approximately 1.47 TB), achieving an aggregate acquisition rate of 6.67 images per second. After architectural optimization, mean latency decreased from 26,344 ms to 126 ms (99.52% reduction), with a median of 117 ms and a 95th percentile of 207 ms. Effective system availability reached 95.29%, with most incidents associated with transient wireless connectivity interruptions. These results demonstrate the feasibility of deploying a distributed thermal IoMT system in a real intensive care unit under hospital information technology governance, offering a preliminary reference model for on-premise, nonintrusive thermal sensing infrastructures in critical care, pending validation across additional institutional and hardware settings.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s10916-026-02455-5.

Introduction

Continuous monitoring is central to critical care. In intensive care units, early detection of physiological deterioration may support timely clinical intervention and improve patient safety [1–3]. Conventional monitoring systems provide high-value information through contact-based sensors and multiparameter devices, but they are usually focused on point measurements and may not capture spatial or surface-level physiological patterns. In critically ill patients, continuous information related to peripheral perfusion, localized inflammation, or temperature distribution may provide complementary information beyond standard monitoring.

Non-contact sensing technologies have gained increasing attention in clinical environments where reducing technological intrusion is important. Infrared thermography is especially relevant because it enables remote acquisition of surface thermal information without attaching sensors to the patient [4–7]. Previous studies have explored thermal imaging for estimating respiratory dynamics, vital signs, and other physiological parameters [8–10]. However, much of this evidence comes from controlled experimental contexts, small prototypes, or algorithm-centered studies.

At the same time, the Internet of Medical Things (IoMT) has expanded the possibilities for distributed physiological data acquisition in healthcare [11, 12]. In theory, IoMT architectures can support scalable, continuous, and automated sensing infrastructures. In practice, deployment in hospitals is constrained by institutional information and communication technology governance, including network segmentation, cybersecurity policies, bandwidth limitations, authentication mechanisms, and data protection requirements.

These constraints are particularly relevant in intensive care units, where new systems must avoid interference with clinical workflows and existing medical infrastructure. Moreover, public cloud-based architectures may be unsuitable or difficult to approve in contexts where local data control, institutional traceability, and compliance with the General Data Protection Regulation are required. Therefore, on-premise architectures remain highly relevant for real-world clinical deployment.

Despite the maturity of many sensing technologies, there remains a gap between experimental IoMT prototypes and operational systems deployed in real hospital environments. The challenge is not only to capture data, but to create systems that are technically stable, scalable, compliant with institutional policies, and compatible with routine clinical activity.

This study addresses that gap by presenting the design, implementation, and operational evaluation of a low-cost, on-premise thermal IoMT system deployed in the intensive care unit of Hospital Universitario HLA Moncloa, Madrid, Spain. The aim was not to clinically validate a specific physiological parameter, but to evaluate whether a distributed thermal monitoring architecture could operate continuously under real hospital conditions using objective operational metrics.

This deployment corresponds to the first of two phases defined in the approved institutional research protocol under which this study was conducted: establishing the technical feasibility, reliability, and governance compliance of the sensing infrastructure, as a precondition for a planned second phase evaluating its clinical utility, including automated detection of patient falls, self-extubation, pressure ulcer development, and cardiac arrest; its role in the differential diagnosis of shock (septic, cardiogenic, or hypovolemic); and monitoring of the time to recovery of normothermia in the postoperative period; together with a structured assessment of ICU staff satisfaction with the system. Establishing this operational infrastructure is a necessary prerequisite before robust clinical validation or AI-assisted thermal analysis can be undertaken in the ICU.

Methods

Study Design

This was a technological implementation and operational evaluation study of a distributed thermal monitoring system deployed in a real intensive care unit. The study focused on system architecture, deployment feasibility, operational stability, latency, data throughput, storage behavior, and scalability. It did not assess diagnostic accuracy, clinical outcomes, or automated decision-support performance.

The project was developed in three phases. First, the architecture was designed and the hardware and software components were selected according to low cost, modularity, scalability, and compatibility with the hospital infrastructure. Second, thermal acquisition nodes were progressively deployed in selected ICU beds and connected to the hospital network under the supervision of the information technology department. Third, system performance was monitored, bottlenecks were identified, and architectural optimizations were implemented and evaluated.

Clinical Setting

The system was deployed in the intensive care unit of Hospital Universitario HLA Moncloa, Madrid, Spain. The ICU has thirteen operational beds. At the time of structured evaluation, seven thermal acquisition nodes were active, with the architecture designed to scale to all thirteen beds.

The deployment used the existing hospital Wi-Fi network, managed by the institutional information technology department. No parallel network infrastructure was created. All image transmission, storage, and visualization were restricted to the internal hospital environment. The backend server was located within the hospital infrastructure and no external cloud service was used for image storage or processing.

System Architecture

The architecture was designed as a distributed environmental sensing system with centralized local storage and visualization. It included four functional layers: acquisition, transmission, storage and processing, and visualization. The overall architecture of the system is shown in Fig. 1.

Fig. 1.

Conceptual architecture of the on-premise thermal IoMT system deployed in the intensive care unit

Each acquisition node consisted of a Raspberry Pi 4B connected to a low-cost radiometric infrared thermal camera with a resolution of 256 × 192 pixels. Nodes were configured to operate autonomously and capture one image per second. Each image was associated with timestamps and technical metadata, including central, maximum, minimum, and mean temperature values. These values were computed over the full thermal frame rather than a patient-segmented region; central temperature corresponded to the fixed geometric center pixel of the image, independent of patient position within the field of view. The thermal acquisition node used in the ICU deployment is shown in Fig. 2.

Fig. 2.

Thermal acquisition node installed in the intensive care unit environment

The acquisition software followed a decoupled model separating image capture, temporary local storage, and transmission. This design allowed image acquisition to continue during temporary network instability. The most recent image was prioritized for near real-time visualization, while pending images were transmitted asynchronously when connectivity was available.

Images were transmitted through the hospital Wi-Fi network to the local server using HTTP within the internal hospital network. Access to the server was restricted to authorized and pre-registered devices. The transmission layer incorporated error handling and retransmission mechanisms to reduce data loss during transient connectivity interruptions.

Device-level access control was enforced through a pre-registered authentication token required on every transmission request, validated against a server-side registry before any data was accepted; requests without a valid token were rejected. Access to the visualization dashboard required individual user authentication with hashed credentials and time-limited session tokens. The Wi-Fi network used for image transmission was logically isolated from the general hospital network, provided and governed by the hospital’s information technology department, with no external internet connectivity, and device admission restricted by MAC-address registration under institutional policy.

The backend adopted a hybrid persistence strategy. Thermal images were stored in a hierarchical file system, while metadata and indexing information were managed through a PostgreSQL database. This separation was intended to support progressive growth of the dataset while maintaining efficient access to recent images and associated metadata.

The visualization layer consisted of a web-based dashboard accessible only within the authorized hospital environment. The dashboard retrieved the latest available image from each node to provide near real-time perception without querying the full historical dataset. At this stage, the dashboard did not provide diagnostic interpretation, automated alarms, or clinical decision support.

A summary of the technical characteristics is shown in Table 1.

Table 1.

Technical characteristics of the deployed thermal monitoring system

Parameter	Description
System type	Distributed environmental thermal monitoring system
Deployment environment	Intensive care unit
Active nodes	7 nodes, scalable to 13 beds
Acquisition device	Raspberry Pi 4B
Thermal sensor	Radiometric infrared camera
Thermal resolution	256 × 192 pixels
Capture frequency	1 image per second per node
Network	Hospital Wi-Fi managed by the information technology department
Transmission	HTTP over local hospital network
Backend	On-premise server
Image storage	Hierarchical file system
Metadata management	PostgreSQL database
Visualization	Local web dashboard
Structured evaluation dataset	5,763,312 thermal images
Storage volume	Approximately 1.47 TB

Data Governance and Ethics

The system was designed according to data minimization and privacy-by-design principles [13, 14]. Thermal images were low-resolution surface thermal maps without optical facial features, audio, or direct patient identifiers. The database stored the bed or node identifier and technical metadata, but no direct clinical identifiers.

Image acquisition was conditional on informed consent and system activation by authorized personnel. If valid consent was not present, image capture could be suspended. Data were stored locally within the hospital infrastructure and were not transferred to external cloud platforms.

The study was conducted in accordance with the Declaration of Helsinki and approved by the Research Ethics Committee for Medicinal Products under approval number CEIm 23/63. Written informed consent was obtained from patients or their legal representatives for the acquisition and use of environmental thermal images for research purposes.

Operational Evaluation

The system was evaluated using technical and operational metrics. End-to-end latency was defined as the time between image capture at the node and availability for visualization on the dashboard. Timestamps were recorded at several points in the data flow, including capture, transmission, reception, and storage.

Latency was summarized using mean, median, 95th percentile, and 99th percentile values. System availability was estimated by comparing the number of images actually produced with the expected number of images for seven nodes operating at approximately one image per second. Connectivity incidents were identified from node telemetry, including Wi-Fi interface drops, reconnections, and prolonged interruptions in image acquisition.

Storage behavior was assessed through cumulative image volume, total storage use, and server resource utilization. Server performance was monitored using CPU and RAM consumption during continuous operation. Scalability was assessed by analyzing whether the system maintained latency, throughput, and resource utilization within acceptable margins during continuous operation with multiple active nodes.

The main structured evaluation was conducted over a continuous 10-day period after system stabilization. In addition, a pre-optimization latency dataset was analyzed to quantify the impact of architectural improvements.

Results

Deployment and System Operation

The system was successfully deployed in a real ICU environment using the existing hospital network infrastructure. Seven nodes operated simultaneously during the structured evaluation period, while the architecture remained scalable to the thirteen operational beds of the unit.

The deployment confirmed the feasibility of integrating low-cost thermal acquisition devices into the clinical environment without creating a parallel network or modifying routine clinical workflows. The system operated as a passive environmental sensing infrastructure and did not require staff to attach additional devices to patients.

During implementation, several practical constraints were identified. The most relevant were network variability, clock synchronization between nodes and server, and power supply instability in some acquisition devices. Clock synchronization was addressed using network time synchronization and a dual timestamp strategy. Power-related instability was mitigated by stabilizing the supply to affected nodes. Network variability was addressed through transmission prioritization, local buffering, and asynchronous retransmission.

Latency Optimization

During the initial deployment phase, the system used a preliminary configuration in which synchronous image transmission and queue accumulation produced high delays. A total of 127,618 latency samples were analyzed in this phase. Mean end-to-end latency was 26,344 ms, median latency was 20,616 ms, and the 95th percentile reached 83,890 ms. These values indicated that images were progressively accumulating in transmission queues, with substantial delays in near real-time visualization.

Architectural optimizations were then introduced. The main changes included decoupling reception from persistence, prioritizing the most recent image for visualization, improving queue management, and adjusting capture and transmission parameters. After these optimizations, 6,917,464 additional latency samples were analyzed.

Mean latency decreased to 126 ms, median latency to 117 ms, and the 95th percentile to 207 ms. This represented a relative mean latency reduction of 99.52%. The improvement affected not only central tendency values but also the upper percentiles, indicating that the optimization eliminated prolonged queue accumulation and stabilized the system under sustained load. The reduction in latency before and after architectural optimization is shown in Fig. 3.

Fig. 3.

Temporal evolution of end-to-end latency before and after architectural optimization

The cumulative distribution of latency values showed a complete shift from multi-second delays to sub-second performance after optimization. This behavior confirmed that the optimized architecture was compatible with near real-time dashboard visualization. The cumulative distribution of latency before and after optimization is shown in Fig. 4.

Fig. 4.

Cumulative distribution function of end-to-end latency before and after architectural optimization

Structured Evaluation After Stabilization

After optimization, the system underwent structured evaluation over a continuous 10-day period with seven active acquisition nodes. During this period, 5,763,312 thermal images were recorded. The generated dataset occupied approximately 1.47 TB, corresponding to an average image size of 275 KB and approximately 147 GB of data per day.

The aggregate acquisition rate was 6.67 images per second. This value was close to the nominal target of seven images per second, corresponding to seven nodes operating at one image per second. The result indicates that the system maintained almost full acquisition capacity under real ICU operating conditions.

The latency distribution remained stable during the structured evaluation. Median and 95th percentile values did not show progressive deterioration as the number of stored images increased. These results suggest that the hybrid persistence model, combining file-system image storage and database metadata indexing, was adequate for continuous operation over millions of images. The latency distribution during the structured evaluation period is shown in Fig. 5.

Fig. 5.

Cumulative distribution function of end-to-end latency during the structured 10-day evaluation period

Operational Stability

Effective system availability during the 10-day structured evaluation was 95.29%, calculated by comparing actual image production with the nominal expected production. Most nodes maintained acquisition rates close to the expected cadence, although some heterogeneity was observed between devices.

The main incidents detected during the evaluation were transient wireless connectivity losses or local node interruptions. A total of 44 Wi-Fi interface drop or reconnection events were detected through periodic telemetry. The cumulative duration of these events was 106,218 s. The median duration was 600 s, corresponding to the minimum observable interval given the telemetry sampling period, and the maximum recorded event lasted 40,807 s.

Analysis of capture continuity showed that some nodes had longer interruptions than others. BOX7 and BOX8 showed the largest temporal discontinuities, while BOX9 and BOX11 showed the most consistent behavior. Importantly, the observed incidents were not associated with progressive degradation of the central server or storage infrastructure.

Storage Growth and Server Performance

The system accumulated 1.47 TB of image data during the structured evaluation. Data growth was approximately linear, consistent with continuous acquisition from multiple nodes. Despite this growth, server resource use remained low throughout the period.

Mean CPU usage was 0.66%, with a 95th percentile of 1.0%. Mean RAM usage was 6.93%, with a 95th percentile of 7.8%. These values indicate that image reception, storage, indexing, and dashboard visualization remained well below the capacity of the central server (see Fig. S01 CPU usage over time: raspberry Pi nodes and server” available as supplementary file S01).

The absence of progressive latency deterioration during storage growth suggests that the architecture can support continuous operation with multiple nodes and millions of images without immediate central infrastructure saturation. This result is relevant for future expansion to the thirteen beds of the unit. The relationship between latency stability and cumulative data growth is shown in Fig. 6.

Fig. 6.

Temporal evolution of latency and accumulated image volume during continuous system operation

Additional operational metrics, including per-node acquisition variability and system resource utilization, are provided in the Supplementary Material (Online Resource 1 and Online Resource 2).

Scalability Projection

The operational results support the feasibility of scaling the system from seven active nodes to full ICU coverage. Server CPU and RAM utilization remained low, and latency did not increase with cumulative data growth during the evaluated period. Monitoring of the embedded nodes also showed low computational load, indicating that image acquisition, compression, and transmission did not saturate the Raspberry Pi devices (see Fig. S02. Hourly Acquisition per Node, available as supplementary file S02).

These findings suggest that the current bottlenecks are more likely to be related to wireless connectivity and local node reliability than to central computation or storage capacity. The modular architecture also provides a basis for future incorporation of edge processing, which could reduce network traffic by performing preprocessing or event detection at the node level.

Discussion

This study demonstrates that a distributed thermal IoMT system can be deployed and operated in a real intensive care unit under existing hospital information technology governance. The main contribution is not the introduction of a new thermal sensor, but the operational demonstration of a complete sensing infrastructure capable of continuous acquisition, local storage, near real-time visualization, and scalable operation within a governed hospital environment.

The results show that the architecture achieved a substantial improvement in latency after optimization. Initial delays of more than 20 s in median latency were reduced to approximately 117 ms. This change transformed the system from a delayed acquisition platform into a near real-time monitoring infrastructure. The improvement was achieved through architectural changes rather than through replacement of the sensing hardware, highlighting the importance of data flow design, queue management, and separation between acquisition, transmission, persistence, and visualization.

The system also maintained stable operation during the structured 10-day evaluation. More than 5.7 million images were generated and stored without evidence of central server saturation. CPU and memory use remained low, and latency did not progressively deteriorate as the dataset grew. These findings support the suitability of a hybrid persistence strategy in which raw images are stored in a hierarchical file system while metadata are indexed in a relational database.

A relevant finding was that the main operational incidents were associated with wireless connectivity and local node interruptions rather than central backend failure. This is consistent with the realities of hospital environments, where research systems must coexist with institutional network policies, variable traffic, scheduled infrastructure maintenance, and security restrictions. Unlike laboratory settings, hospital deployments require adaptation to conditions that are not fully controlled by the research team.

From a medical systems perspective, this deployment highlights the importance of designing IoMT infrastructures around governance constraints from the beginning. Data protection requirements, local storage, access control, and institutional supervision were not secondary administrative issues but structural design drivers. The use of an on-premise architecture enabled local control of data and avoided dependence on external cloud services, which may be difficult to approve in sensitive clinical environments.

This differs from existing clinical video-distribution infrastructures, such as surgical-video information systems or ward surveillance cameras, which typically require post-hoc de-identification steps to protect patient privacy, such as blurring or blanking applied to video streams before release [15, 16]. In the system reported here, non-identifiability is a structural property of the acquisition modality itself: the low-resolution radiometric thermal sensor does not resolve the optical facial features required for identification, rather than relying on a downstream processing step that can be omitted, misconfigured, or reversed. The technical contribution of this work is therefore not video or image transmission as such, but the design, empirical operational characterization, and governance-compliant deployment of a low-cost, on-premise, thermal-specific IoMT architecture.

The system described here should be interpreted as a foundational infrastructure rather than a clinical decision-support tool. As introduced above, this corresponds to the first of two phases defined in the approved research protocol under which this study was conducted, with the present study establishing technical feasibility, reliability, and governance compliance ahead of the planned clinical-utility phase. At this stage, the system does not generate alarms, classify patient status, or provide diagnostic recommendations. Its value lies in enabling continuous acquisition of structured thermal datasets under real clinical conditions, intended to support the event-detection and pattern-recognition components planned for that second phase.

The study also provides practical lessons for similar deployments. First, small-scale progressive implementation is preferable to immediate full-unit deployment. Second, close coordination with hospital information technology teams is essential. Third, local buffering and asynchronous transmission are necessary for resilience in shared wireless environments. Fourth, system evaluation should include operational metrics such as latency, availability, connectivity incidents, storage growth, and resource utilization, not only sensor accuracy.

Several limitations must be acknowledged. This was a single-center deployment, and results may vary in hospitals with different network infrastructure, security policies, or ICU layouts. Results are also specific to the hardware configuration used (Raspberry Pi 4B acquisition nodes with a low-cost 256 × 192 radiometric thermal sensor); performance may differ with other embedded platforms or thermal cameras. The system was not clinically validated against reference physiological measurements, and no diagnostic or prognostic claims can be made from the current study. The evaluation period, although sufficient to assess operational feasibility after stabilization, does not replace long-term multicenter validation. Similarly, scalability to the full thirteen-bed configuration was projected from operational trends observed with seven active nodes rather than directly tested, and actual performance at full capacity remains to be empirically verified; scalability, infrastructure cost, and security are widely recognized as open challenges for edge computing in healthcare more broadly [17]. In addition, the system remains a pre-product research infrastructure and has not undergone certification as a medical device.

Additional limitations relate to security and measurement scope. Transport between acquisition nodes and the backend used HTTP without TLS encryption; this risk was mitigated by the network-layer isolation described in Methods rather than by transport-layer encryption, and no formal security audit or penetration testing of the deployed infrastructure was performed as part of this study. As this represents a pilot-phase deployment, transport-layer encryption was deprioritized relative to establishing operational feasibility and is identified as a security-hardening step for any future wider or production deployment. Reported temperature metrics (maximum, minimum, mean, and central values) were computed over the entire thermal frame rather than a patient-isolated region. As the field of view could include non-patient elements (e.g., medical equipment) and patients could be partially covered by bedding, these values should not be interpreted as validated measurements of patient body temperature; isolating a patient-specific region of interest was outside the scope of this operational evaluation and is identified as future work.

Future work should address multicenter deployment, longer monitoring periods, integration of edge computing, and clinical validation of specific thermal indicators. Particular attention should be given to privacy impact assessment, algorithm governance, model monitoring, and human oversight if automated analysis is incorporated. The current architecture provides a basis for this evolution because it separates acquisition, storage, visualization, and future analytical components.

Conclusions

This study demonstrates the feasibility of deploying and operating a distributed, on-premise thermal IoMT system in a real intensive care unit under hospital information technology governance. The system achieved near real-time performance after architectural optimization, sustained continuous acquisition from seven nodes, and generated more than 5.7 million thermal images without central infrastructure degradation.

The proposed architecture provides an early-stage architectural template for nonintrusive environmental sensing in critical care, derived from a single-center, short-term deployment using a single hardware configuration; broader generalization would require validation in other institutional settings. Its main value lies in bridging the gap between experimental thermal imaging prototypes and operational hospital systems. Although further clinical validation is required, the deployed infrastructure establishes a foundation for future edge computing, automated thermal analysis, and data-driven monitoring applications in intensive care.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

E.A. and J.F. contributed equally to this work. They conceived the system architecture, led technological development, coordinated and emplemented the deployment of the monitoring system. S.G. provided the clinical perspective, supervised integration within the intensive care environment, and supported ethical and regulatory alignment. J.B. contributed to conceptualization, scientific framing, project supervision, and manuscript revision. All authors reviewed and approved the final manuscript.

Funding

This work was funded by the Fundación ASISA–Universidad Europea de Madrid Chair in Health Sciences, within the TermoCam_V1 and TermoCam_V2 projects, 2024-GC01. The funding supported the acquisition of equipment and materials required for development and deployment of the thermal monitoring system. The funding body had no role in study design, data analysis, interpretation of results, or manuscript preparation.

Data Availability

No datasets were generated or analysed during the current study.

Code Availability

The custom code used for system acquisition, transmission, storage, and visualization is not publicly available at this stage because it is part of an ongoing institutional research infrastructure. Access may be considered by the corresponding author upon reasonable request and subject to institutional approval.

Declarations

Ethics approval

This study was conducted in accordance with the Declaration of Helsinki and was approved by the Research Ethics Committee for Medicinal Products under approval number CEIm 23/63. Clinical trial number: not applicable.

Consent to participate

Written informed consent was obtained from patients or their legal representatives for the acquisition and use of environmental thermal images for research purposes.

Consent to publish

No directly identifiable participant images or personal data are included in this manuscript. The thermal images generated by the system do not contain optical facial features.

Use of artificial intelligence

AI-assisted tools were used to support language editing, structural refinement, and formatting of the manuscript. No AI tool was used for data acquisition, statistical analysis, image manipulation, reference fabrication, or autonomous scientific interpretation. The authors reviewed and verified all content and take full responsibility for the integrity and accuracy of the manuscript.

Competing interests

The authors declare no competing interests.