Japan divides acute inpatient care into roughly 335 secondary medical care areas, which serve as planning units under the 8th National Health Care Plan. Conventional regional comparisons often rely on single risk-adjusted metrics, which can conflate patient-mix differences with true structural differences in intensive care units (ICUs). The authors present MedZone Embedder, a framework to derive data-driven indicators of regional structural vulnerability by learning a similarity space of secondary medical care areas from a national ICU outcomes registry.
This approach reframes regional assessment away from one-dimensional rankings toward a representation that answers: which areas are structurally similar, and which are structural outliers exhibiting potential institutional deficits.
Each secondary medical care area is represented by a feature vector formed from aggregated intensive care provision indicators derived from the Japan Intensive Care Patient Database (JIPAD). Concrete indicator groups included in the representation are:
The representation also incorporates within-area variance of these indicators to capture internal heterogeneity. The method intentionally omits hierarchical processing by facility type, producing area-level aggregates rather than facility-stratified profiles.
The principal representation-learning pathway is a contrastive autoencoder with a multilayer perceptron encoder architecture (32 → 16 → 8) and a symmetric decoder. The model is trained in a self-supervised fashion using an objective that combines reconstruction loss and normalized temperature cross-entropy (NT-Xent) applied to noise-augmented views of the same input.
Training produces an 8-dimensional embedding for each area. This low-dimensional embedding space is designed to support cosine-similarity-based searches for similar areas and to make structural relationships between areas explicit.
To identify structural outliers and quantify regional vulnerability, the framework applies anomaly-scoring methods in the embedding space. Proposed scorers include:
Raw anomaly scores from these methods are normalized to a unified vulnerability score on a 0–1 scale, facilitating ranking and interpretation across areas.
Recognizing operational constraints, the authors provide a deterministic fallback using principal component analysis (PCA). This PCA pathway is intended for settings where deep-learning libraries are unavailable or when the number of areas is small, situations in which contrastive learning may be impractical.
The framework was implemented and deployed within an operational ICU decision support system on a managed cloud platform. For a proof of concept (PoC), the authors used synthetic facility-level aggregate data constructed to follow the JIPAD indicator schema; no real registry data were accessed at this stage. The PoC focused on five secondary medical care areas in Kyoto Prefecture.
The deployed system produced several outputs:
The authors report that at the small scale of the PoC, the contrastive autoencoder defaulted to the PCA projection. The deep-learning implementation has been completed and unit-tested, but training and evaluation on actual registry data remain pending data-use approval and expanded data integration.
Using the synthetic PoC dataset, the system generated profiles and an embedding that supported similar-area lookups and vulnerability ranking. The vulnerability ranking identified areas characterized by low patient numbers, low case diversity, and poor aggregated outcomes. The PoC demonstrated feasibility of the overall pipeline and the practical outputs that stakeholders could review within an operational decision support environment.
The paper emphasizes that these PoC results are not empirical findings from registry data but demonstrations of the framework and software behavior on synthetic data matching the JIPAD schema.
The authors pre-specify a staged validation program:
The paper notes that training and evaluation on actual JIPAD registry data are contingent on obtaining data-use approval and completing broader data integration.
The PoC used only synthetic, facility-level aggregate data; Kyoto University's ethics committee waived ethical approval for this stage. The authors state that subsequent validation stages involving real registries will be subject to full institutional review processes.
Competing interests are declared: one author represents a company that retains commercial rights for non-ICU applications, and the other chairs a collaboration that has royalty-free access for ICU-domain use. The deployment within ICU services is licensed to the named network under those terms.
MedZone Embedder proposes a data-driven framework to represent the structural characteristics of Japanese secondary medical care areas using JIPAD-derived indicators. By producing an embedding space and normalized vulnerability scores, the framework aims to distinguish areas that are structurally similar from those that are outliers, offering an alternative to single-indicator regional comparisons. The contribution is presented as a method and deployed implementation with a pre-specified validation plan; national empirical application depends on future registry access and validation stages.