Clinical Pathways as Safety Specifications for Physical AI in Hospital Wards

AI summary

arXiv: 2607.19827v1 [cs.RO], 22 Jul 2026 — Gabriele Franchini, Giulio Mallardi, Michele De Carolis, Filippo Lanubile (Computer Science Dept., University of Bari, Bari, Italy)

Overview

Research area: Safe Physical AI, runtime safety monitoring, healthcare robotics, and trustworthy AI — specifically the use of Clinical Pathways as executable safety specifications for embodied systems in hospital wards.

Technical level: Advanced. The paper assumes familiarity with Signal Temporal Logic (STL), runtime verification, conformal prediction, and robotic middleware concepts; it is conceptual and architectural rather than empirical.

Scope (one sentence): The paper proposes a conceptual, not-yet-deployed robotic architecture in which a Runtime Safety Monitor hosted on an assistive robot checks multimodal patient and device signals against Clinical Pathway constraints encoded in temporal logic, producing interpretable, severity- and confidence-graded safety alerts.

What This Paper Is About

Physical AI systems such as assistive robots, wearable sensors, and smart medical devices increasingly operate in hospital wards alongside vulnerable patients and clinical staff, where safety depends on whether the intended care process is actually followed despite sensor noise, communication failures, uncertainty, and misuse. The authors argue that statistical anomaly detection is insufficient for this setting because an anomaly score does not say which clinical requirement was violated, whether the deviation is clinically meaningful, or who should intervene. Their goal is to reinterpret Clinical Pathways — which already encode medication schedules, measurement windows, physiological targets, and dependencies between clinical actions — as explicit runtime safety specifications that an embodied system can verify.

Key Contributions

1. A reformulation of Clinical Pathways as runtime-verifiable safety specifications for embodied medical AI, shifting the guiding question from "is this signal anomalous?" to "is the prescribed care being delivered safely?"
2. A robotic architecture that integrates wearable sensing, smart medical devices, and embodied monitoring on an assistive robot, extending beyond the static edge-node home-care solution of earlier work.
3. A Runtime Safety Monitor (RSM) design combining temporal prediction, uncertainty-aware reasoning (via conformal prediction), and constraint-based verification to produce interpretable safety-violation reports.
4. A taxonomy of three safety-event classes — physiological deviations, system and embodiment failures, and adversarial tampering — handled by a single monitor through the same machinery.

Main Findings

• Clinical Pathways supply the missing runtime semantics: CPs already encode bounded physiological ranges, scheduled windows, ordered actions, and cross-event dependencies; the paper supplies what is missing by encoding them in a temporal logic for continuous signals, yielding a full specification denoted φ_CP.

• Four constraint types are identified in a clinician-authored CP: (i) range constraints on physiological quantities, (ii) temporal constraints on timing and frequency of measurements, (iii) sequence constraints on the order of clinical actions, and (iv) cross-modal constraints linking events across devices (for example, a pill-dispenser opening must be followed within a bounded delay by a measurable physiological effect).

• Two STL operators carry a direct clinical reading: the always operator □φ, used for example so that every scheduled blood-pressure measurement falls within the target band, and the bounded eventually operator ◇_[Δ1, Δ2]φ, used so that a medication intake is followed by a measurable effect within the prescribed pharmacological window.

• Robust semantics enable graded alerts: each formula returns a real-valued robustness σ where positive values indicate satisfaction with a margin and negative values quantify severity of a violation, allowing alerts to be ranked by severity rather than issued as binary alarms.

• Severity and confidence are deliberately separated: the RSM emits a tuple ⟨τ, κ, σ, ρ, ξ⟩ where σ quantifies how badly the prescribed care is violated and ρ indicates how confidently a violation can be claimed — two orthogonal axes that the paper states a single anomaly score would conflate.

• Alerts are intrinsically explainable: each alert points, via ξ, to the offending sub-formula of φ_CP, without post-hoc attribution, and an event manager maps each tuple to a safety action such as notifying the clinician, locking the dispenser, or aborting a procedure.

• Governance is preserved by design: the translation from clinician-authored CP to STL is performed offline in the edge cloud and distributed to the robot at session start; the robot only evaluates specifications it cannot author or modify.

• Illustrative trace (explicitly not an empirical evaluation): in the hypertension scenario, the CP prescribes two antihypertensive pills (07:00, 21:00) with expected blood-pressure response within 1–3 hours, five daily blood-pressure measurements with target band BP ∈ [110, 140] mmHg, and counter constraints requiring five measurements and two intakes over 24 hours. Table I reports one day in which one event of each class arises: 07:00 pill (dispenser open), σ = +18, ρ = 0.97, no violation; 08:55 bp = 152 mmHg, σ = −12, ρ = 0.94, violated sub-formula φ_BP, class C1; 15:00 bp_meas (9 s window), σ = −21, ρ = 0.71, violated sub-formula φ_HW, class C2; 21:00 pill (scheduled), σ = +15, ρ = 0.96, no violation; 21:02 pill (unscheduled), σ = −8, ρ = 0.42, violated sub-formula φ_DOSE, class C3.

• The three event classes are distinguished by which sub-formula fails: C1 physiological deviations (for example hypertensive spikes, post-medication trajectories that fail to converge within the prescribed delay, missed scheduled measurements); C2 system and embodiment failures (battery depletion, probe detachment, Bluetooth errors, and extensions to localization drift, manipulation failures during pillbox interaction, and missed deadlines on safety-critical topics); C3 adversarial tampering (data injection on the wearable bus, spoofing of patient presence to the robot's perception stack, or physical manipulation of the smart pill dispenser to forge adherence events).

Methodology in Plain English

The authors present a conceptual reference architecture organized into three cooperating subsystems — sensing, embodied, and edge cloud. The sensing subsystem provides raw evidence: patient-worn Bluetooth devices (sphygmomanometer, ECG patch, pulse oximeter, smart pill dispenser) capture physiological and adherence signals prescribed by the CP, ambient sensors supply context, and on-robot perception (RGB-D camera, microphone array) provides an independent channel for presence verification so that a tampered wearable can be cross-checked against the robot's own observations. The embodied subsystem runs on the assistive robot — a platform such as TIAGo Pro is named as a possible instantiation — and exposes navigation, human-robot interaction, device control, and sensor fusion as lifecycle-managed middleware nodes whose operational states are observable to the monitor, helping disambiguate sensor faults from clinical events. The Runtime Safety Monitor sits at the conceptual core and runs a typed pipeline: a multivariate time-series learner predicts physiological trajectories; a constraint checker evaluates the STL formulae online and returns a real-valued robustness σ rather than a Boolean verdict; and a calibrated uncertainty subsystem based on conformal prediction wraps the predictor with prediction intervals that provide distribution-free coverage under exchangeability assumptions, so violations are flagged only when the uncertainty envelope exits the safe region. Dataflow combines pull-based acquisition (the CP schedules sensor acquisition within clinically relevant windows) with push-based alerts that propagate locally to clinicians and nursing staff without traversing the cloud. The architecture is described as middleware-agnostic, with ROS 2 given as a reference implementation rather than a requirement. The authors validate the idea only through an illustrative one-day scenario trace, which they label as not an empirical evaluation.

Why This Matters

The paper argues for a broader shift in runtime safety for embodied AI: from data-driven anomaly detection that treats deviations as statistical outliers, to specification-based monitoring where deviations are interpreted against an explicit, human-authored model of intended behavior. In clinical settings this is argued to be especially natural because the specification already exists in the form of the Clinical Pathway — what has been missing is the machinery to make it a runtime artifact. The paper frames this as a contribution to Safe Physical AI that operationalizes domain-specific clinical knowledge and assists nursing staff.

Real-world applications (as framed by the paper):

• Hospital wards where assistive robots, wearable sensors, and smart medical devices operate alongside vulnerable patients and clinical staff.
• Hypertension monitoring with scheduled antihypertensive pills, timed blood-pressure measurements, and dosage counters (the running example).
• Medication adherence supervision, including detecting a pill-dispenser event that lacks the expected downstream physiological response and locking the dispenser.
• Infrastructure-fault detection, distinguishing hardware and robot-platform failures from genuine clinical events.

Industry relevance: The architecture is aimed at robotics and healthcare technology developers building middleware-agnostic embodied systems — ROS 2 is named as a reference implementation and TIAGo Pro as a possible platform — and at clinical governance, since the robot evaluates but cannot author or modify the specifications it checks.

Future Directions

• Specification authoring: translating a clinician-authored CP into STL presupposes either a template library or a clinician-facing authoring tool, and parameter values instantiating these templates must be grounded in clinical guidelines and pharmacokinetic data; STL expressiveness ultimately bounds the safety properties one can state.
• Distribution shift: conformal coverage guarantees rely on exchangeability assumptions that physiological signals routinely violate around clinical events, and adaptive variants mitigate but do not eliminate the issue.
• Clinical validation: the pseudo-trace establishes feasibility, not efficacy; the architecture is not deployed, and studies with real devices, patients, and clinicians are required before any deployment claim.
• Extensions: patient-specific calibration to adapt CP parameters to individual baselines while preserving the logical structure of the specification, offline retrospective checking of authored CPs against patient history for clinical governance, and empirical studies of how clinicians and nursing staff interpret RSM outputs, especially when severity and confidence diverge.

Target Audience

Researchers and practitioners in Safe Physical AI, runtime verification, and healthcare robotics; roboticists building embodied systems for clinical environments; developers working with ROS 2 and assistive robot platforms; and clinical informatics or governance specialists interested in turning Clinical Pathways into enforceable runtime safety contracts. The paper is best suited to readers with some background in temporal logic and runtime monitoring, since it assumes those concepts rather than introducing them.

Funding note: The authors state the research was co-funded by the Complementary National Plan PNC-I.1 ("DARE", PNC0000002, CUP: B53C22006420001). The paper reports no empirical evaluation and no deployed system.

Authors’ abstract

Ensuring safety in Physical AI systems operating in real-world environments is a critical challenge, particularly in hospital wards where vulnerable patients, clinical staff, medical devices, and assistive robots coexist. In this paper, we reinterpret Clinical Pathways as explicit runtime safety specifications for embodied medical AI. We propose a conceptual robotic architecture that integrates wearable sensors, smart medical devices, and assistive robotic components into a unified framework for real-time safety monitoring. At its core, a Runtime Safety Monitor (RSM) evaluates multimodal physiological and system-level signals against clinically defined constraints derived from the prescribed care process. Rather than relying solely on statistical anomaly detection, the proposed approach combines temporal prediction, uncertainty-aware reasoning, and constraint-based verification to identify safety violations. The RSM targets three classes of events: physiological deviations from prescribed care, hardware and communication failures, and potential data tampering or misuse. This work contributes to Safe Physical AI by operationalizing domain-specific clinical knowledge as enforceable safety constraints, bridging learning-based perception and runtime safety monitoring to assist nursing staff in real-world hospital wards.