{"schema_version":1,"research_id":"eoa_inverse_innovation_exp04_external_evaluation_20260802","source_assessment_id":"invariant_mode_decomposition_design__medicine_healthcare:PROPOSAL_FIRST:v0","cell_id":"invariant_mode_decomposition_design__medicine_healthcare","search_queries":["dynamic mode decomposition patient deterioration vital signs hospital ward","spectral analysis multivariate vital signs early warning clinical deterioration eigenvector transition matrix","site:ahrq.gov patient deterioration failure to rescue hospital ward early warning systems","Royal College Physicians NEWS2 deterioration standard official","\"dynamic mode decomposition\" \"vital signs\" patient","\"dynamic mode decomposition\" ICU patient","transition matrix eigenvalues patient deterioration multivariate vital signs","Koopman operator patient deterioration ICU physiological","\"Dynamic Modeling of Patient Vital Signs\" transition matrix title authors","site:dl.acm.org \"Dynamic Modeling of Patient Vital Signs\"","\"Dynamic Modeling of Patient Vital Signs\" deterioration","clinical deterioration dynamic transition matrix vital signs temporal graph model","site:fda.gov clinical decision support software final guidance 2022 non-device recommendations healthcare professional independently review basis","site:healthit.gov predictive decision support intervention source attributes HTI-1 transparency risk management","FDA device software functions clinical decision support patient deterioration alert guidance","NICE evidence standards framework digital health technologies early warning algorithm","eCART early warning score multicenter prospective validation JAMA Network Open clinical deterioration 2020","electronic Cardiac Arrest Risk Triage eCART ward deterioration primary study","Advance Alert Monitor mortality implementation study NEJM Kaiser 2020 deterioration","TREWS sepsis prospective implementation clinical deterioration alert study","site:pubmed.ncbi.nlm.nih.gov \"The value of vital sign trends for detecting clinical deterioration on the wards\"","site:pubmed.ncbi.nlm.nih.gov \"Comparison of the Between the Flags\" eCART","site:pubmed.ncbi.nlm.nih.gov \"Automated Identification of Adults at Risk for In-Hospital Clinical Deterioration\"","site:pubmed.ncbi.nlm.nih.gov eCARTv5 prospective validation"],"sources":[{"source_id":"S1","title":"Surveillance Monitoring to Improve Patient Safety in Acute Hospital Care Units","publisher":"Agency for Healthcare Research and Quality Patient Safety Network","url":"https://psnet.ahrq.gov/perspective/surveillance-monitoring-improve-patient-safety-acute-hospital-care-units","source_class":"GOVERNMENT_OR_REGULATOR","publication_date":"2023-04-26","accessed_at":"2026-08-02","claims_supported":["Failure to recognize inpatient deterioration is a material patient-safety problem.","AHRQ reports estimates attributing 10%–13% of hospital deaths to inaction or failure to recognize deterioration.","General-ward vital signs are commonly intermittent, and static or intermittent scores may miss unwitnessed deterioration.","Monitoring must be adapted to ward workflow and alert communication constraints."]},{"source_id":"S2","title":"National Early Warning Score (NEWS) 2","publisher":"Royal College of Physicians","url":"https://www.rcp.ac.uk/resources/national-early-warning-score-news-2/","source_class":"OFFICIAL_GUIDANCE","publication_date":"2017-12-19","accessed_at":"2026-08-02","claims_supported":["NEWS2 is an established aggregate physiological score for standardized recognition and response to acute illness.","NEWS2 uses six routinely measured physiological parameters plus supplemental oxygen.","NHS England and NHS Improvement formally endorsed NEWS2 for identifying acutely ill hospital patients.","NEWS2 is intended to supplement clinical judgment and has defined escalation guidance.","The binary oxygen supplement component can fail to reflect a rapid increase in oxygen requirement, illustrating a known coordinate/aggregation limitation."]},{"source_id":"S3","title":"Advance Alert Monitor Program: An Automated Early Warning System for Adults At Risk for In-Hospital Clinical Deterioration","publisher":"Agency for Healthcare Research and Quality Patient Safety Network","url":"https://psnet.ahrq.gov/innovation/advance-alert-monitor-program-automated-early-warning-system-adults-risk-hospital","source_class":"GOVERNMENT_OR_REGULATOR","publication_date":"2023-06-14","accessed_at":"2026-08-02","claims_supported":["Kaiser Permanente Northern California is an identifiable adopter of automated deterioration monitoring and implemented AAM in 21 hospitals.","AAM combines predictive analytics, EHR integration, virtual nurse review, rapid-response workflow, and clinician decision authority.","Deployment requires multidisciplinary clinical, IT, monitoring, rapid-response, and continuing-review resources.","A deployed end-to-end deterioration-monitoring workflow is established prior art."]},{"source_id":"S4","title":"Multicenter Development and Prospective Validation of eCARTv5: A Gradient-Boosted Machine-Learning Early Warning Score","publisher":"Society of Critical Care Medicine / Critical Care Explorations","url":"https://pubmed.ncbi.nlm.nih.gov/40151509/","source_class":"PRIMARY_RESEARCH","publication_date":"2025","accessed_at":"2026-08-02","claims_supported":["eCARTv5 is a close supervised-risk comparator using demographics, vital signs, documentation, and laboratory values to predict ICU transfer or death within 24 hours.","The model was developed across seven hospitals and validated retrospectively and prospectively across 21 hospitals.","Reported retrospective AUROC was 0.834 versus 0.766 for NEWS and 0.704 for MEWS.","The validation supported FDA clearance, showing a high evidentiary bar for a clinical deterioration product."]},{"source_id":"S5","title":"The Value of Vital Sign Trends for Detecting Clinical Deterioration on the Wards","publisher":"Resuscitation / Elsevier","url":"https://pubmed.ncbi.nlm.nih.gov/26898412/","source_class":"PRIMARY_RESEARCH","publication_date":"2016-05","accessed_at":"2026-08-02","claims_supported":["Coordinate-level temporal trends are an established comparator to static scores.","In 269,999 admissions, adding trends improved AUROC from 0.74 to 0.78.","Simple change from the immediately preceding value worsened average AUROC, showing that temporal feature definition matters.","The study directly supports temporal holdout comparison against simple trend baselines."]},{"source_id":"S6","title":"Dynamic Modeling of Patient Vital Signs: Leveraging Markov Chain Principles with Neural Networks for Irregular Time-Series Prediction","publisher":"University of Utah author manuscript repository","url":"https://users.cs.utah.edu/~tch/publications/pub334.pdf","source_class":"PRIMARY_RESEARCH","publication_date":"2024","accessed_at":"2026-08-02","claims_supported":["Patient vital-sign vectors have already been represented as states in a dynamic system with learned state transitions.","The work explicitly addresses irregular multivariate observations and cross-variable interactions.","It forecasts state evolution with a neural transition function rather than decomposing a fitted linear operator into stable and unstable modes.","The manuscript is an anonymous submission and does not establish clinical effectiveness or adoption."]},{"source_id":"S7","title":"Clinical Decision Support Software: Guidance for Industry and Food and Drug Administration Staff","publisher":"U.S. Food and Drug Administration","url":"https://www.fda.gov/media/109618/download","source_class":"OFFICIAL_GUIDANCE","publication_date":"2026-01-29","accessed_at":"2026-08-02","claims_supported":["FDA's current guidance distinguishes non-device CDS from device software functions using four statutory criteria.","Non-device CDS must enable a healthcare professional to independently review the basis of a recommendation and avoid primary reliance on it.","Software analyzing signals from a signal-acquisition system for a medical purpose may remain a device function.","A patient-specific deterioration alert therefore requires intended-use and regulatory classification analysis before live deployment; labeling it a reassessment prompt is not by itself dispositive."]},{"source_id":"S8","title":"Decision Support Interventions: §170.315(b)(11) Test Method and Clarifications","publisher":"Assistant Secretary for Technology Policy / Office of the National Coordinator for Health Information Technology","url":"https://healthit.gov/test-method/decision-support-interventions/","source_class":"GOVERNMENT_OR_REGULATOR","publication_date":"2024","accessed_at":"2026-08-02","claims_supported":["Predictive decision support includes models producing prediction, classification, recommendation, evaluation, or analysis from training data.","Certified health IT must support source attributes, privileged configuration, and access to plain-language model information.","For supplied predictive interventions, intervention risk management covers validity, reliability, robustness, fairness, intelligibility, safety, security, privacy, mitigation, and governance.","The framework calls for ongoing validation, fairness monitoring, feedback capability, and governance over data acquisition and use."]}],"problem_evidence":{"support":"STRONG","rationale":"Government patient-safety material identifies missed deterioration as consequential and reports that static/intermittent monitoring can fail; NEWS2, eCART, and AAM demonstrate that hospitals already devote substantial infrastructure to the problem. Evidence supports the broad problem, but not the narrower assertion that a reproducible amplifying eigenmode is a common cause of missed ward deterioration.","source_ids":["S1","S2","S3","S4","S5"]},"stakeholder_evidence":{"support":"STRONG","rationale":"Kaiser Permanente implemented AAM across 21 hospitals with clinical, informatics, virtual-nurse, and rapid-response participation, while NHS bodies formally endorsed NEWS2. These identify credible adopters and authorizers with expressed demand for earlier recognition. No source expresses demand specifically for spectral-mode explanations.","source_ids":["S2","S3"]},"prior_art":{"proximity":"ADJACENT_PRIOR_ART","closest_analogues":[{"name":"NEWS2","similarity":"Combines multiple routinely collected physiological measurements into an interpretable ward deterioration trigger tied to clinician reassessment and escalation.","remaining_difference":"It is a static aggregate threshold score and does not estimate a transition operator, classify joint dynamics by gain, or suspend output on spectral-gap, eigenvector-drift, and reconstruction failures.","source_ids":["S2"]},{"name":"eCARTv5","similarity":"Uses multivariate vital signs, laboratory values, and other EHR data to generate repeatedly updated 24-hour ward-deterioration risk and has large multicenter prospective validation.","remaining_difference":"It directly predicts risk using gradient-boosted trees rather than reporting invariant coupled directions, their decay or amplification, spectral separation, and modal residual fidelity.","source_ids":["S4"]},{"name":"Advance Alert Monitor","similarity":"Provides an operational precedent for automated EHR deterioration surveillance, centralized nurse review, rapid-response communication, clinician authority, and continuing performance review.","remaining_difference":"Its product is a high-risk score embedded in an extensive response program, not a locally linear modal state representation with gap/drift suppression.","source_ids":["S3"]},{"name":"Coordinate-level vital-sign trend models","similarity":"Use temporal changes in ward vital signs and improve deterioration discrimination over current values alone.","remaining_difference":"They model slopes, extrema, or other individual-variable trends rather than coupled invariant directions and their scalar gains.","source_ids":["S5"]},{"name":"Neural Markov-style patient-state transition modeling","similarity":"Treats the multivariate vital-sign vector as a patient state and learns repeated transitions, including cross-variable interactions and irregular sampling.","remaining_difference":"It forecasts states through a neural transition function and does not establish eigendecomposed deterioration modes, spectral safeguards, clinical alert performance, or deployed workflow.","source_ids":["S6"]}],"distinctive_claim_remaining":"On a temporally held-out adult-ward dataset, a frozen locally linear transition model can yield sufficiently conditioned, spectrally separated, temporally stable, and reconstructively faithful coupled modes; at a matched alert burden, activity in prespecified persistent or amplifying modes provides reproducible incremental warning or lead time beyond NEWS2/eCART-style risk and coordinate-trend comparators, while its original-variable loadings remain coherently interpretable. This is contrastive and falsifiable but presently unverified.","confidence":"HIGH"},"implementation_evidence":{"support":"MODERATE","rationale":"The component technologies—multivariate EHR extraction, repeated state modeling, retrospective silent evaluation, EHR alerts, nurse review, and model governance—are feasible and precedented. However, no located source validates eigendecomposed ward-transition modes. Irregular and informative sampling, treatment-confounded transitions, missing urine/mental-status/support data, ill-conditioned or non-normal operators, local nonstationarity, subgroup errors, and alert burden remain material technical and safety risks. Offline read-only work is feasible under local approval; live display requires workflow testing, model governance, privacy/security review, and an FDA intended-use determination.","source_ids":["S3","S4","S5","S6","S7","S8"]},"scores":{"meaningful_impact":{"score":4,"rationale":"Earlier recognition targets severe outcomes including arrest, ICU transfer, and death, but the candidate's realized benefit is unmeasured.","source_ids":["S1","S3"]},"stakeholder_pull":{"score":4,"rationale":"Hospitals and national health authorities demonstrably adopt deterioration systems, although no source requests a modal method.","source_ids":["S2","S3"]},"incremental_advantage":{"score":2,"rationale":"Traceable dynamic modes and explicit validity suppression are plausible advantages, but strong static, trend, and supervised-risk comparators already exist and no incremental result is available.","source_ids":["S2","S4","S5"]},"distinctiveness_plausibility":{"score":3,"rationale":"No close ward-deployment analogue with eigenmode gain, spectral-gap monitoring, and reconstruction gating was found, but dynamic state-transition and multivariate temporal modeling are established adjacent art.","source_ids":["S4","S5","S6"]},"technical_implementability":{"score":3,"rationale":"An offline prototype is technically plausible, but irregular sampling, intervention confounding, missingness, non-normality, near-degeneracy, and short local validity windows could defeat the proposed interpretation.","source_ids":["S5","S6"]},"adoption_authority_feasibility":{"score":3,"rationale":"Hospital safety and model-governance leaders are credible authorizers, but live use would require local workflow approval and unresolved FDA classification analysis.","source_ids":["S3","S7","S8"]},"evidence_readiness":{"score":3,"rationale":"A bounded retrospective evaluation can use existing EHR data, but availability and reliability of all proposed state variables and sufficient deterioration events at one ward are not verified.","source_ids":["S4","S5","S6"]},"safety_net_benefit":{"score":4,"rationale":"A non-autonomous reassessment prompt with automatic suppression on drift, residual, missingness, conditioning, and spectral-gap failures could add a useful safety layer without removing existing pathways; this remains a design claim.","source_ids":["S7","S8"]},"scalability":{"score":2,"rationale":"Deterioration workflows have scaled across hospitals, but the candidate deliberately requires local fitting, monitoring, and retirement, so cross-ward portability and operational burden are uncertain.","source_ids":["S3","S4"]}},"score_confidence":"MODERATE","costs":{"first_evidence":{"band_2026_usd":"50K_TO_250K","scope":"Preregistered one-ward retrospective study: data extraction and harmonization, transition-model implementation, temporal validation, three comparator implementations, subgroup analysis, two-clinician review, governance, and reporting.","confidence":"MODERATE","assumptions":["Existing EHR extracts are accessible under local approval.","Approximately 0.8–1.5 data-scientist FTE-years plus part-time informatics, clinical, biostatistical, and governance effort.","No live EHR integration, patient contact, regulatory submission, or new sensing.","Resource-equivalent estimate, not a vendor quote."],"source_ids":["S3","S4","S5","S6"]},"initial_deployment_startup":{"band_2026_usd":"250K_TO_1M","scope":"Silent prospective integration on one ward, production data pipeline, identity and access controls, logging, validation dashboards, human-factors design, cybersecurity/privacy review, quality-system documentation, and regulatory assessment.","confidence":"LOW","assumptions":["Existing EHR integration interfaces can be reused.","Output remains hidden from treating clinicians during startup.","Includes multidisciplinary clinical and IT work but not a pivotal clinical trial or FDA submission fees.","Local integration complexity is unknown."],"source_ids":["S3","S7","S8"]},"operational_launch":{"band_2026_usd":"1M_TO_5M","scope":"Controlled live one-ward launch with prospective safety evaluation, staffed escalation workflow, training, round-the-clock support, audit, human-factors testing, adverse-event review, and any required device-quality or regulatory work.","confidence":"LOW","assumptions":["Hospital retains existing rapid-response capability.","Launch includes sufficient clinical and technical coverage to avoid an unsupported alert channel.","Regulatory pathway and evidence requirements are unresolved and could move cost outside this band.","No hospital-wide rollout or new bedside hardware."],"source_ids":["S3","S7","S8"]},"annual_recurring":{"band_2026_usd":"250K_TO_1M","scope":"One-site production operation: data-quality surveillance, mode/gap/residual and subgroup monitoring, clinical owner time, incident response, retraining or retirement reviews, cybersecurity, software maintenance, and recurring governance.","confidence":"LOW","assumptions":["One hospital or a small group of wards.","Existing EHR and rapid-response staff absorb part of the workflow.","Material mode instability or dedicated 24-hour monitoring staff could exceed the band.","No new continuous-monitoring hardware."],"source_ids":["S3","S8"]}},"verified_pipeline_gates":{"externally_supported_problem":{"status":"YES","reason":"Official and primary sources directly document missed ward deterioration, adverse consequences, and limitations of intermittent or static monitoring.","source_ids":["S1","S2","S3","S5"]},"externally_credible_adopter_or_authorizer":{"status":"YES","reason":"Kaiser Permanente, hospital clinical leadership, NHS bodies, and certified-health-IT governance roles establish credible adopters and authorizers for deterioration decision support.","source_ids":["S2","S3","S8"]},"distinct_testable_incremental_claim":{"status":"YES","reason":"The claim can be tested as stable, reconstructively faithful modal warning with incremental performance or lead time at matched alert burden versus NEWS2/eCART and coordinate trends.","source_ids":["S2","S4","S5"]},"bounded_next_evidence_step":{"status":"YES","reason":"A frozen, temporally split, one-ward retrospective study with named comparators, metrics, safety checks, and no live output is bounded and executable.","source_ids":["S4","S5","S6"]},"no_unresolved_safety_or_authority_stop":{"status":"YES","reason":"No stop prevents the proposed offline read-only study if local data-governance approval is obtained. This does not authorize live use; FDA classification, privacy/security, human factors, and clinical governance remain pre-launch requirements.","source_ids":["S7","S8"]},"credible_cost_scope_and_range":{"status":"UNCERTAIN","reason":"The resource categories are supported by deployed-program and governance precedents, but no local data-access, EHR-integration, staffing, regulatory-pathway, or procurement estimates were available; launch and recurring bands are therefore low-confidence.","source_ids":["S3","S7","S8"]}},"next_evidence_step":"With a hospital data steward and ward patient-safety sponsor, preregister an offline study using 12–18 months of one adult ward for fitting and the immediately following 3–6 months for locked temporal testing. Freeze the state variables, interval, missingness handling, outcome, operator regularization, eigenvector-conditioning ceiling, spectral-gap floor, retained-mode rule, reconstruction budget, subgroup definitions, and alert threshold. Compare NEWS2, the locally available eCART or strongest supervised score, and prespecified coordinate-slope/extrema models. At one matched alert rate per 100 patient-days, measure AUPRC, AUROC, calibration, decision-curve net benefit, median warning time, sensitivity, residual error and structure, eigenvector conditioning, spectral separation, mode rotation, and subgroup error. Two blinded clinicians should independently review no more than 100 de-identified trajectories for explanation coherence and reification risk. Falsify progression if the candidate does not improve AUPRC or median lead time over the best comparator at matched burden; if calibration is materially worse; if modes rotate, become ill-conditioned or lack the frozen spectral gap during ordinary operation; if clinically consequential structured residuals exceed budget; if any key subgroup has a prespecified unacceptable sensitivity or false-alert disparity; or if fewer than 70% of reviewed explanations are judged coherent. No output should enter live care.","blocking_evidence":["No external evidence shows that adult-ward deterioration is commonly organized into stable, spectrally separated amplifying modes.","No patient-level data verify that the proposed variables are captured at a sufficiently regular and reliable cadence on the target ward.","No held-out comparison establishes incremental warning, calibration, net benefit, or alert burden versus NEWS2, eCART, and simple temporal features.","No evidence establishes acceptable operator conditioning, spectral gaps, reconstruction fidelity, drift, or subgroup performance.","No clinician study shows that mode-to-variable explanations are consistently understandable without reifying modes as disease entities.","Live-use FDA status, privacy/security controls, human-factors safety, operational ownership, and local cost remain unresolved."],"research_disposition":"PARTNERED_RESEARCH_PROGRAM","world_novelty_boundary":"The search found established static scores, supervised multivariate deterioration models, operational alert programs, coordinate-trend models, and a neural patient-state transition model. It did not find a directly matching adult-ward product or study combining a fitted local transition operator, eigenmode gain classification, original-variable loadings, spectral-gap and mode-drift suppression, and held-out reconstruction gating. This is only a bounded search result; world novelty, patentability, freedom to operate, market size, and realized impact remain unmeasured.","arm":"PROPOSAL_FIRST","candidate_version":0,"controller_recommendation":{"action":"STOP_EMPIRICAL_RESEARCH_NEEDED","repairable":true,"material_progress_observed":true,"progress_targets":["Obtain a one-ward data-access determination and verify availability, cadence, semantics, missingness, and intervention timestamps for every proposed state variable.","Preregister the locked temporal study, event definition, comparators, matched alert burden, minimal incremental effect, conditioning/gap/drift/residual thresholds, subgroup limits, and falsifiers.","Run the offline temporal validation and blinded explanation review; report negative and subgroup results as well as aggregate discrimination.","Seek a documented FDA intended-use classification and local privacy, security, clinical-model-governance, and human-factors assessment before any live display.","Replace broad deployment cost bands with local EHR integration, staffing, regulatory, monitoring, and support estimates."],"reason":"Bounded web research verifies a serious problem, credible adopters, established rivals, feasible offline workflow, and a contrastive modal claim. It cannot determine whether stable and useful deterioration modes exist or outperform strong comparators; that requires proprietary patient-level data and empirical testing. Live authority and cost details also require institution-specific work."}}