{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp04_retrieval_first_paired20_20260802","cell_id":"invariant_mode_decomposition_design__medicine_healthcare","arm":"PROPOSAL_FIRST","candidate_id":"imdd_mh_coupled_deterioration_mode_watch_v0","hypothesis_id":null,"version":0,"title":"Coupled Deterioration Mode Watch for Adult Inpatient Wards","problem":"During ward monitoring, clinicians may see several modest changes in vital signs, organ-function measurements, and support requirements without recognizing that their combination is evolving along a self-reinforcing physiologic direction. Coordinate-level thresholds and static aggregate scores can obscure whether the joint state is decaying toward baseline, persisting, oscillating, or amplifying across observation intervals.","actors":["adult inpatients on a participating ward","bedside nurses","ward physicians and advanced practice clinicians","rapid-response clinicians","clinical informatics staff","hospital patient-safety leadership"],"observable_state":"At each fixed observation interval, a patient is represented by a standardized state vector containing available heart rate, respiratory rate, blood pressure, oxygen saturation, oxygen-support level, temperature, urine output, mental-status indicator, and selected laboratory measurements. The observable warning state is a rising projection onto a locally estimated mode whose gain indicates persistence or amplification, accompanied by a traceable loading pattern such as concurrent circulatory, respiratory, and renal movement that may not cross individual alert thresholds.","consequence":"If a growing coupled direction is missed, clinical reassessment and escalation may occur only after conspicuous coordinate-level abnormalities emerge; if a spurious mode is treated as meaningful, patients and staff may instead be exposed to unnecessary alarms, testing, or escalation.","affected_objective":"Support timely, proportionate recognition of evolving inpatient deterioration while limiting uninformative alerts and preserving clinician control over diagnosis and treatment.","intervention":"Build a non-autonomous mode-watch layer over one ward's existing observation workflow. Estimate a locally valid transition operator from consecutive patient-state vectors, decompose it into modes and gains, and project each new state change into that basis. Display only modes meeting prespecified persistence, consequence, spectral-separation, and residual-fidelity rules. Each display shows the contributing original variables, direction of change, operating window, residual warning, and a prompt for clinician reassessment under existing protocols; it does not diagnose a condition or select treatment. Suspend output when the basis drifts, the spectral gap is inadequate, missingness exceeds the declared scope, or reconstruction residuals breach tolerance.","structural_mapping":[{"archetype_element":"Transformation Scope","domain_realization":"The transformation is the fixed-interval update from one multivariate inpatient physiologic state to the next within a declared ward population and care regime."},{"archetype_element":"State-Vector Definition","domain_realization":"Standardized vital signs, support levels, organ-function measurements, mental-status indicators, and missingness flags form the original-coordinate state."},{"archetype_element":"Invariant Mode Basis","domain_realization":"Eigenvectors of the fitted local transition operator represent combinations of measurements that approximately preserve their direction across successive updates."},{"archetype_element":"Modal Gain Spectrum","domain_realization":"Each eigenvalue characterizes whether its associated physiologic combination is damped, persistent, oscillatory, or growing within the fitted window."},{"archetype_element":"Stable/Unstable Mode Partition","domain_realization":"Modes are separated by prespecified gain boundaries into decaying, marginal, oscillatory, and potentially amplifying monitoring classes."},{"archetype_element":"Modal Intervention Map","domain_realization":"For each actionable mode, the system maps its original-variable loadings to an existing clinician reassessment pathway, without asserting a diagnosis or treatment effect."},{"archetype_element":"Reconstruction Residual Check","domain_realization":"The retained modes must reconstruct held-out state transitions within a declared tolerance, with structured residuals reviewed for omitted clinically important behavior."},{"archetype_element":"Mode Drift Monitor","domain_realization":"Rolling comparisons detect changes in mode direction, ordering, gain, and spectral separation after shifts in population, measurement, or care practice."},{"archetype_element":"Interpretation Scope Contract","domain_realization":"Every output states that the decomposition is local, approximate, observational, and invalid outside the named population, interval, variables, and care regime."}],"mechanism_mapping":[{"mechanism_slug":"eigendecomposition_workflow","role":"Factor the explicit fitted transition matrix into a complete set of physiologic modes and scalar gains, while reporting eigenvector conditioning and complex or near-degenerate modes.","counterfactual_removal":"Without decomposition, the layer reverts to coordinate-level coefficients and cannot distinguish invariant coupled directions or classify their repeated evolution."},{"mechanism_slug":"modal_stability_analysis","role":"Classify each fitted mode as decaying, persistent, oscillatory, or amplifying inside the declared local window, making growing directions candidates for review.","counterfactual_removal":"Without stability classification, mode amplitude alone cannot distinguish a large resolving pattern from a smaller pattern that is projected to grow."},{"mechanism_slug":"modal_sensitivity_sweep","role":"Perturb modal coordinates in the fitted model and measure changes in the prespecified deterioration outcome, while logging cross-mode movement and treating results as local associations rather than causal treatment effects.","counterfactual_removal":"Without the sweep, modes would be ranked mainly by gain or current magnitude, which may elevate visually dominant but outcome-insensitive directions and hide coupling."},{"mechanism_slug":"residual_reconstruction_test","role":"Determine how many modes may be retained by testing reconstruction error and residual structure on temporally held-out patient transitions.","counterfactual_removal":"Without residual testing, compression could discard a low-amplitude but clinically consequential pattern without making the omission visible."},{"mechanism_slug":"spectral_gap_monitor","role":"Track separation between retained and discarded modes and rotation of retained directions, suppressing output when modal identity is unreliable.","counterfactual_removal":"Without gap and drift monitoring, near-degenerate or changing modes could be presented as stable clinical entities after the decomposition has lost its warrant."},{"mechanism_slug":"spectral_decomposition_report","role":"Translate each retained mode back to named measurements and document permitted interpretation, couplings, residuals, validation window, and retirement conditions.","counterfactual_removal":"Without the report, clinicians could mistake a mathematical mode for an independent disease process or a causal treatment target."}],"causal_chain":["Consecutive multivariate ward observations define an explicit local state-transition operator rather than a collection of independent thresholds.","Decomposition identifies combinations of measurements that the fitted transition approximately preserves as directions and assigns each a scalar gain.","Stability analysis distinguishes damped combinations from persistent, oscillatory, or amplifying combinations.","Outcome sensitivity and consequence rules select review-relevant modes instead of selecting solely by variance, magnitude, or gain.","New patient observations are projected into the retained modes, and an alert candidate arises only when a relevant mode is active and passes fidelity and spectral-separation checks.","The interface translates that mode back into its contributing measurements and prompts an authorized clinician to reassess the patient using existing clinical pathways.","Residual, drift, and spectral-gap failures suppress the modal interpretation and trigger model review rather than silent extrapolation."],"baseline":"Ordinary practice is review of individual trends and threshold breaches, optionally summarized by a static aggregate early-warning score, followed by clinician judgment. This baseline is directly available, interpretable, and not dependent on a locally stationary transition model.","nearest_rivals":["A static aggregate early-warning score is the strongest rival because it combines multiple measurements, is easier to operationalize, and may capture all decision-relevant signal without estimating temporal modes.","Single-variable trend and threshold alerts can provide clearer reasons for review and may be more robust when coupled dynamics are unstable.","A supervised time-series risk model can estimate near-term deterioration directly and may outperform a constrained modal representation, although its internal state need not correspond to invariant directions."],"remaining_contrastive_claim":"Conditional on adequate local stability, spectral separation, and held-out reconstruction, the candidate represents deterioration as repeated evolution along traceable coupled directions and monitors the validity of those directions; the strongest static-score rival combines coordinates into risk but does not explicitly classify their joint update as damped, persistent, oscillatory, or amplifying.","authority_safety":{"decision_authority":"The responsible bedside clinician retains authority over assessment, diagnosis, testing, escalation, and treatment; patient-safety leadership and the designated clinical model-governance group control deployment, suspension, and retirement.","authorized_first_step":"Clinical informatics staff may conduct an offline, read-only evaluation on de-identified historical data from one adult inpatient ward under local data-governance approval; no score or mode is shown to treating staff and no care pathway is changed.","excluded_actions":["autonomous diagnosis","autonomous medication, fluid, oxygen, transfer, or testing orders","suppression of existing alarms or escalation pathways","use outside the declared ward population or operating window","presentation of modal sensitivity as a causal treatment effect","patient-level output when residual, drift, missingness, conditioning, or spectral-gap safeguards fail"],"halt_rollback":"Suppress all candidate output and revert to the unchanged existing monitoring workflow if data definitions change, eigenvectors are ill-conditioned or unstable, the retained/discarded spectral gap crosses its prespecified minimum, held-out residuals exceed tolerance, subgroup error review identifies unacceptable concentration, or governance approval expires. Because the first step is offline and read-only, rollback consists of deleting derived evaluation outputs according to the approved retention plan and making no clinical deployment."},"negative_tests":{"strongest_counterevidence":"On temporally held-out data, a prespecified static aggregate score or simple coordinate-trend baseline matches or exceeds the modal candidate on discrimination, calibration, alert timeliness, and alert burden, while the candidate adds unstable modes, structured residuals, or poor subgroup performance.","problem_falsifier":"The inferred problem is falsified for the scoped ward if adjudicated deterioration episodes are already preceded by clear individual threshold breaches or static-score escalation, and coupled modal amplitude provides no reproducible incremental warning before those signals.","intervention_falsifier":"The intervention is falsified for progression if its retained modes fail temporal validation, rotate materially across ordinary operating periods, lack adequate spectral separation, cannot reconstruct clinically consequential transitions within the preset budget, or cannot be translated into consistent clinician interpretations and reassessment decisions during a later approved simulation.","risks":["Treatment actions may alter subsequent measurements, causing the fitted operator to mix physiology with clinician response and making modal gains non-causal.","Irregular sampling, imputation, and informative missingness may create artificial modes.","A locally linear model may fail during abrupt deterioration or after a large intervention.","Non-normal or nearly degenerate operators may yield fragile eigenvectors despite apparently stable eigenvalues.","Modes may encode documentation, device, or ward-practice artifacts rather than patient physiology.","Performance or residual failures may concentrate in patient subgroups masked by aggregate evaluation.","A modal display may create automation bias or encourage staff to reify a mode as a disease entity.","Additional alerts may increase workload even when their mathematical signal is reproducible."]},"next_evidence_step":"Using a preregistered analysis plan, fit the transition operator on an earlier time block from one adult ward and freeze the state definition, observation interval, mode-selection rules, residual budget, spectral-gap minimum, and comparator thresholds. Evaluate on the immediately subsequent held-out block against the existing static score and coordinate-trend baseline for calibration, alert timing, alert burden, reconstruction residuals, mode conditioning and drift, and subgroup error patterns. Have two blinded clinicians review a bounded sample of de-identified trajectories only to judge whether mode-to-variable explanations are coherent; do not expose outputs in live care or estimate treatment effects.","prior_art_status":"UNSEARCHED","revision_record":{"parent_version":null,"progress_targets_addressed":[],"conceptual_changes":["Initial proposal instantiates invariant-mode decomposition as a locally bounded monitor of coupled inpatient physiologic transitions."],"operational_changes":["Initial version limits the first step to one ward, historical de-identified data, frozen rules, existing comparators, and no clinical display or care change."],"evidence_changes":[],"claim_changes":["No novelty, prevalence, demand, or effect-size claim is made; the contrastive claim is conditional and prior art remains unsearched."]}}