KIRETT: AI in Rescue Operations

• KIRETT is a wrist-worn decision-support system integrating AI, knowledge graphs, and real-time vital sign monitoring for emergency rescue operations.
• It employs an artificial neural network trained on extensive rescue records to perform dynamic situation recognition and provide actionable treatment guidance.
• The platform connects wearable sensors, middleware integration, and human feedback to optimize time-critical interventions and improve patient outcomes.

KIRETT, abbreviated from “Artificial Intelligence in Rescue Operations,” is a wrist-worn decision-support system for rescue personnel that combines artificial intelligence for situation recognition, knowledge-graph-based treatment guidance, and real-time integration of patient vital signs. Its stated purpose is to improve first aid during rescue operations by providing contextual recommendations for actions and operations, minimizing incorrect medical treatments, increasing the probability of patient survival, and reducing the risk of long-term impairments. Across the current publications, KIRETT is presented as a wearable platform that links field data acquisition, graph-based clinical workflow modeling, and human-confirmed treatment recommendations for time-critical emergency care (Zenkert et al., 29 Sep 2025, Nadeem et al., 11 Aug 2025).

1. Project scope and clinical rationale

KIRETT is motivated by the observation that rescue operations are increasingly frequent and are conducted under severe time pressure. In such settings, first responders must provide personalized and optimized healthcare in the shortest possible time while estimating the patient’s condition from freshly recorded vital data. The project literature emphasizes that, in emergencies, treatment needs to be fast, accurate, and patient-specific, while environmental obstacles, medical difficulties, and abrupt changes in health vitals can force paramedics to shift treatment during ongoing care (Nadeem et al., 11 Aug 2025).

The project therefore targets a specific gap: rescue personnel may not be able to fully grasp or operationalize all relevant procedural knowledge during a time-dependent intervention. KIRETT is intended to make on-the-spot calculated, evaluated, and processed knowledge available during treatment. Initial use-cases explicitly mentioned in the project include Acute Coronary Syndrome (ACS), pre-hospital trauma life support, and hypoglycemia management. The overall framing is not autonomous care delivery, but operational support for first responders and medical experts under acute constraints (Zenkert et al., 29 Sep 2025).

2. System architecture and knowledge representation

The architecture described for KIRETT consists of several interlinked modules: a wrist-worn wearable device, a Knowledge Graph (KG), a Situation Detection (SD) module, middleware, a graphical user interface (GUI), and a messenger system based on a message queue infrastructure. The middleware mediates communication between the wearable, the KG, external medical devices such as the ZOLL X-Series, and the central KIRETT database that stores real-time vital data. The GUI presents recommendations, queries, and alerts to first responders and requires active confirmation when human decisions are necessary, thereby adding a safety layer for both patients and responders (Nadeem et al., 11 Aug 2025).

The KG is the central knowledge representation. It is constructed from the official rescue operation manual and is defined as a directed, cyclic, weakly connected graph,

$KG = \{V,E\},$

where $V$ denotes nodes and $E$ denotes edges. The encoded graph contains 37 BPRs, 39 SAAs, 3046 nodes, and 4467 relations. Node categories include StartNode and StopNode, BPRNode and SAANode, JumpNode, DecisionNode (YN/OR), ProcedureNode and InvasiveProcedureNode, ActionNode, and DisplayNode and WarningNode. Edge categories include $R^n$ for priority-ordered workflow links, yes/no branches, association links, and additionalInformation links. Node properties include id, name, d_type, value, and min/max, enabling dynamic queries and data-driven logic.

The graph construction process is described as iterative and expert-driven. After initial text mining and manual encoding from the rescue manual, the KG underwent regular review by medical professionals for correctness and usability. The graph is explicitly not complete in the graph-theoretic sense that all nodes are directly connected; this is done to preserve semantic correctness and procedural order rather than maximize connectivity. Neo4j is used as the graph database, and Graphlytics is used for visualization (Nadeem et al., 11 Aug 2025).

3. Situation detection, vital signs, and recommendation logic

KIRETT combines symbolic treatment-path traversal with statistical situation recognition. The AI core is an Artificial Neural Network (ANN) trained on over 300,000 real rescue records from the past 5 years. Input features include patient vitals such as SpO$_2$, blood pressure, and blood sugar, together with diagnostic information and text entries from rescue personnel. Categorical variables are processed with one-hot encoding, while unstructured medical reports and descriptions are processed using TF-IDF vectorization. The ANN outputs a probability vector over complication groups rather than specific diseases,

$\vec{p} = [p_1, p_2, \ldots, p_n],$

with categories including cardiovascular, CNS, pulmonary, respiratory, abdominal, psychiatric, metabolic, gynecologic-obstetrical, infections, and special diseases. The wearable continuously collects new data and reruns the neural network to update situation recognition in real time (Zenkert et al., 29 Sep 2025).

The vital-sign integration layer is structurally central rather than ancillary. The project reports that 78% of treatment paths and 51% of standard procedures in expert manuals rely on at least one vital sign for branching or decision-making. Middleware retrieves measurements from peripheral medical devices via Bluetooth, stores them in a database, and serves them to the graph component as needed. At a relevant DecisionNode, the system requests the current value, compares it against threshold properties encoded in the KG, and either proceeds automatically when safe or solicits responder confirmation. The recommendation logic is thus a combination of graph traversal, threshold-based branching, and situation pre-recognition (Nadeem et al., 30 Sep 2025).

A representative subset of the manually analyzed guideline dependencies is summarized below.

Vital sign	No. of occurrences in guidelines	Data source
Age	42	Database Entry
Systolic Blood Pressure	34	Measurements
Weight	13	Measurements
SpO2 (Oxygen Saturation)	13	Measurements
Heart Frequency	13	Measurements
Blood Glucose	11	Measurements
Pain Scale / Temperature / ETCO2 / etc.	10 or fewer	Database Entry / Measurements

Within the GUI, relevant values are displayed with timestamps, “Unknown” labels indicate missing data, and thresholds can drive color-coded alerts, with green indicating an in-range value and red indicating an out-of-bounds value. The system also supports automated suggestions for treatment-path skipping when a vital-value condition unambiguously determines the next step, while preserving user override and accept/decline interactions. The published workflow therefore integrates knowledge-graph traversal, rule-based thresholding, and ANN-based pre-recognition into a unified bedside-like rescue interface (Nadeem et al., 30 Sep 2025).

4. User interaction and operational workflow

Operationally, KIRETT is designed around a wrist-worn touch interface with constrained but high-salience interaction patterns. The described input channels include language selection, acquisition of time-series vital signs from devices via WiFi or Bluetooth, information from the control center, and touchscreen input from paramedics through question-and-answer dialogs. The touchscreen output includes AI-based situation recognition, classification of patient severity, specialized treatment plans, and recommended actions. The hardware demonstrator includes a 3D-printed case and additional modules for power and connectivity, and the interface is explicitly optimized for clarity, simplicity, and wearable constraints (Nadeem et al., 29 Sep 2025).

The qualitative evaluation identifies several recurring UX themes. Operators judged the treatment paths medically accurate and closely aligned with the official manual. Half of participants praised the interface for simplicity, clear text, and large buttons, and the majority reportedly understood the UI quickly, including participants with lower prior familiarity with digital technology. At the same time, some participants requested more contextual overview beyond strictly step-by-step guidance, arguing for situational awareness and adaptation during field care. Additional requests included more intuitive navigation, smartphone-like scrolling, color schemes suitable for colorblind users, and haptic or visual feedback for confirmed interaction.

Voice interaction was discussed as a desirable but contested modality. Operators expressed interest in hands-free control, but also pointed to ambient noise, patient comfort, and privacy as substantial barriers during emergencies. Many preferred a multimodal design in which touch and voice coexist. Further requested features included integrated pediatric protocols, a timer for time-critical actions, automatic medication dosing calculators, and improved documentation support. Environmental suitability also emerged as a requirement: the device must remain readable and operable under varying lighting and weather conditions and must be easy to disinfect (Nadeem et al., 29 Sep 2025).

5. Empirical evaluation

The quantitative evaluation reports a two-day user study with 14 rescue professionals at the Fire Station in Siegen, Germany, including participants from both Fire Station Siegen and the German Red Cross, Siegen. The distribution was 10 participants (71%) from Fire Station Siegen and 4 participants (29%) from the German Red Cross, Siegen. Experience levels ranged from trainee (29%) to more than 11 years (29%). Participants first completed a short survey, then spent 90 minutes with the testing device and performed 2–3 standardized rescue test-cases per team, followed by a 23-question post-use survey. Data collection used LimeSurvey, and analysis used MAXQDA and Microsoft Excel (Nadeem et al., 30 Sep 2025).

The reported quantitative results show strong perceived need for digital support. Digitalization in rescue operations was rated “very relevant” by 86% and “relevant” by 7%, for a combined 93%. The importance of AI/ML in rescue work was rated “important” by 57% and “very important” by 43%, yielding 100% agreement on the importance of AI/ML integration. At the same time, 57% reported little personal experience with AI and 35% reported no experience, indicating a gap between perceived relevance and practical familiarity.

The same study highlights operational constraints of the current prototypes. Hygiene was a dominant requirement: cleaning and disinfecting the device was rated “very important” by 100% of participants, and interaction with gloves was rated “important” by 79%. Real-time monitoring of vitals was rated “very important” by 64% and “important” by 15%, while connectivity to Electronic Health Records or a National EPR was rated “very important” by 57% and “important” by 21%. Integration of vitals and user actions for hospital reporting was rated “very important” by 79%.

Ergonomics were weaker than functional desirability. For the LCD case, 36% rated it “very uncomfortable” and 29% “uncomfortable,” while the remaining 50% rated it comfortable or very comfortable. For the FPGA case, 50% rated it “very uncomfortable” and 29% “uncomfortable.” Weight was also criticized: 43% rated the device “very heavy” and 43% “heavy.” The paper characterizes these findings as descriptive rather than inferential, reflecting the pilot nature of the study and the small sample size (Nadeem et al., 30 Sep 2025).

6. Limitations, contested issues, and projected development

The current KIRETT publications identify several technical and organizational limitations. A central tension is flexibility versus rigidity: step-by-step automation is valued as a reminder system that can prevent omissions under stress, but operators also note that practical deviations from protocol are sometimes necessary. Data reliability is another recurrent concern. The system depends on accurate and current external data, especially vital signs, and operators raised the possibility of sensor malfunction or data-transmission failure. Separate work on vital-sign integration also identifies Bluetooth connectivity, device interoperability, and real-time data completeness as technical challenges in dynamic rescue settings (Nadeem et al., 29 Sep 2025, Nadeem et al., 30 Sep 2025).

The AI layer is also treated cautiously. Bayesian-network-based suggestions informed by historical rescue data are described as potentially valuable for surfacing alternative treatments or overlooked scenarios, but some operators feared that such recommendations might introduce new biases or reinforce existing ones. Age-related digital literacy was raised as a possible acceptance issue, with younger operators described as more receptive than some older paramedics. Voice control likewise remains ambivalent because of ambient noise, patient perception, and privacy constraints. These issues situate KIRETT within a broader class of human-in-the-loop clinical support systems rather than fully automated medical agents.

Recommended next steps are concrete and extensive. The publications call for broader contextual overviews of treatment paths, accessibility improvements such as colorblind-friendly palettes and scalable fonts, multimodal touch and robust voice control with noise filtering, integration of pediatric workflows, multi-language support, timers and medication dosing calculators, haptic and visual feedback, stronger documentation functions, and hardware refinement toward lighter, slimmer, more ergonomic, disinfectable, and durable devices. At the system level, planned directions include more extensive field testing with broader cohorts, continuous user involvement in development, further research on bias, enhanced explainability and transparency for auditability and regulatory compliance, continued expansion of the KG to additional pathologies and triage logic, and scalability to mass-casualty events, including a stated potential synergy with digital twin concepts for system-level patient and resource monitoring (Nadeem et al., 29 Sep 2025, Nadeem et al., 11 Aug 2025).