Clinical AI Therapeutic Monitoring Through Phase-Shift Detection

Regulatory framework

The regulatory environment for clinical therapeutic monitoring is denser than the surface vocabulary suggests. HIPAA establishes the privacy and security baseline for any system that ingests behavioral, physiological, or self-report data linked to an identifiable patient, with breach notification, minimum-necessary disclosure, and access-control obligations that shape system architecture from the data-ingestion layer upward. FDA 21 CFR 820 governs the quality system that must surround any device-grade software, and 21 CFR Part 11 governs electronic records and electronic signatures, requiring tamper-evident storage, audit trails, and validated state transitions for any clinical record the system produces.

The FDA AI/ML SaMD framework, articulated in the 2021 action plan and the subsequent predetermined change control plan guidance, establishes the regulatory expectation that adaptive algorithms used in clinical decision support are characterized by their intended use, predicate equivalence where applicable, and a rigorously bounded change protocol. IEC 62304 specifies the software lifecycle obligations, including risk classification, hazard analysis, and verification activities that must accompany any therapeutic monitoring software intended for clinical use. ISO 13485 wraps these obligations in a quality management system that auditors expect to see operationalized in the daily engineering process, not produced retrospectively.

The EU MDR raises the bar further, classifying software that informs therapeutic decisions as a medical device with notified-body oversight, post-market surveillance, and clinical evaluation obligations. ANSI/AAMI EC57 supplies the algorithmic-performance methodology that monitoring systems are increasingly held to, requiring validated sensitivity, specificity, and predictive value characterizations on representative datasets. IEEE 11073 PHD defines the device-side interoperability surface that allows physiological telemetry to enter the monitoring stack without bespoke integration. USCDI v3 and FHIR R4 define the clinical-data interoperability surface that allows monitoring outputs to be exchanged with EHRs, payers, and care-coordination platforms without semantic loss. The cumulative effect is a stack in which a therapeutic monitoring capability must be simultaneously a privacy-compliant data system, a validated medical device, a regulated AI/ML artifact, and an interoperable clinical-record producer.

Architectural requirement

The architectural implication is that therapeutic monitoring must be built around a defined inference object that is privacy-bounded, lifecycle-managed, performance-characterized, and interoperably representable. The inference object must capture clinical state at a temporal resolution that matches the dynamics of the conditions being monitored, which in psychiatric and behavioral therapeutics means days and hours, not weeks. It must be derivable from heterogeneous inputs, including session content, digital phenotyping streams, and structured self-report, without privileging any single modality. It must be expressible in FHIR-compliant resources for exchange and in EC57-compliant performance terms for validation. And it must be governed by a change-control regime compatible with the FDA AI/ML predetermined change control framework.

Few existing therapeutic monitoring stacks meet these requirements simultaneously. Symptom inventories are interoperable but lack temporal resolution. Digital phenotyping streams have temporal resolution but lack a structured clinical inference object. Clinical assessments are validated but episodic. The architectural gap is the absence of a continuous, structured, validated inference layer that unifies these inputs under a clinically meaningful state representation.

Why procedural compliance fails

Most therapeutic monitoring systems satisfy regulatory obligations procedurally rather than architecturally. PHQ-9 and GAD-7 are administered every two to four weeks, results are stored in the EHR under HIPAA-compliant access controls, and the resulting record is treated as the monitoring artifact. This posture fails at three levels.

First, the temporal granularity is mismatched to the dynamics of the conditions being monitored. Mood and anxiety disorders produce coherence phase shifts on timescales of days, not weeks. A patient deteriorating between assessments is invisible to the system until the next scheduled instrument, by which time the phase shift may have stabilized at a worse baseline or precipitated a crisis. The procedural artifact is regulator-facing, not patient-facing.

Second, symptom inventories measure surface phenomena, not the structural dynamics that EC57 expects monitoring algorithms to characterize. Two patients with identical PHQ-9 scores can have radically different coherence trajectories, one stabilizing with residual symptoms and one deteriorating with emergent symptoms. Symptom-score monitoring cannot discriminate these cases because it does not model the underlying state space. Under the FDA AI/ML framework, this is the kind of underspecified algorithm that increasingly fails to clear the validation bar for therapeutic decision support.

Third, digital phenotyping deployments that have attempted to close the temporal gap have generally done so without a structured inference object. They surface raw behavioral metrics, which are neither interoperably exchangeable as clinical findings under USCDI v3 nor validatable as diagnostic outputs under EC57. The clinician receives a dashboard that is rich in data and poor in clinical meaning, and the procedural compliance posture cannot be elevated to architectural compliance simply by adding more streams.

What the AQ primitive provides

The Adaptive Query disruption-modeling primitive supplies the structured clinical inference object that the regulatory stack expects. For each patient, it maintains a continuously updated estimate of position on the promotion-containment continuum, where promoted states are flexible, adaptive, and coherent, and contained states are rigid, defensive, and compensatory. The estimate integrates session content, digital phenotyping signals, and structured self-report under a single inference model whose change protocol is compatible with the FDA AI/ML predetermined change control framework.

The primitive resolves clinical state across a five-axis diagnostic: attention coherence, emotional regulation, relational stability, narrative consistency, and behavioral flexibility. This vector representation captures imbalanced deterioration, in which a patient appears stable on aggregate symptom inventories while deteriorating on a specific axis that predicts adverse outcomes. The vector is the inference object that EC57 performance characterizations apply to and that FHIR R4 resources represent in clinical exchange.

Phase-shift detection is the primitive's central clinical output. Rather than producing only an instantaneous state estimate, it produces a forward-looking inference of imminent transitions between promoted and contained regimes, identifying the patient trajectories that warrant clinician attention before symptoms cross instrument-detection thresholds. The detection logic supports therapeutic dosing guidance: contained-regime patients receive supportive interventions that reduce containment pressure, while promoted-regime patients can tolerate challenging interventions that consolidate gains. The primitive thereby converts continuous monitoring from a passive surveillance posture into an active dosing-calibration substrate, which is the use profile that the FDA AI/ML SaMD framework most directly anticipates.

Compliance mapping

The primitive maps onto the regulatory stack at every layer. Under HIPAA, the inference architecture supports minimum-necessary processing and granular access control, because the inference object is a clinically actionable summary rather than a raw stream of behavioral exhaust. Under 21 CFR 820 and 21 CFR Part 11, the primitive's deterministic state-update logic and immutable audit trail satisfy electronic-record and quality-system expectations. Under IEC 62304 and ISO 13485, the inference object is a defined software item with a documented lifecycle, hazard analysis, and verification regime.

Under the FDA AI/ML SaMD framework, the primitive is presented with a predetermined change control plan that bounds adaptive updates within a clinically validated envelope, satisfying the most challenging regulatory expectation in the contemporary AI/ML clinical software space. Under ANSI/AAMI EC57, the five-axis vector and the phase-shift detector are characterized with sensitivity, specificity, and predictive-value statistics on representative datasets, producing the algorithmic-performance evidence that auditors and notified bodies expect. Under EU MDR, the same evidence package supports clinical evaluation and post-market surveillance documentation. Under IEEE 11073 PHD, the primitive consumes physiological telemetry through standardized interfaces, and under USCDI v3 / FHIR R4 it emits monitoring findings as standard clinical resources, ensuring that the inference object circulates through EHRs, care-coordination platforms, and patient-facing tools without semantic loss.

Adoption pathway

Adoption proceeds in three architecturally distinct stages. The first stage is inference-object instantiation: the organization configures the primitive against its existing therapeutic data flows, ingesting session metadata, digital phenotyping streams where available, and structured self-report instruments. The output is a baseline coherence trajectory for each enrolled patient, established under the existing HIPAA covered-entity and business-associate framework without new data-collection authority.

The second stage is clinical workflow integration. Trajectory views and phase-shift alerts surface in the clinician interface, with the diagnostic specificity necessary to drive therapeutic dosing decisions. The procedural instruments remain in place as confirmatory measures, and the EC57-style performance characterization of the primitive's outputs is established against the existing record. For teletherapy platforms, this stage closes the between-session monitoring gap that procedural systems cannot. For psychiatric inpatient settings, it provides the early-warning capability that crisis-prevention workflows have lacked.

The third stage is regulatory and interoperability alignment. The primitive's outputs are emitted as FHIR R4 resources for exchange under USCDI v3, the change-control plan is filed under the FDA AI/ML SaMD framework where applicable, and the EU MDR clinical evaluation file is maintained for European deployments. The inference object becomes the organization's defended therapeutic monitoring capability, satisfying every layer of the regulatory stack through the same architectural mechanism that produces the clinical benefit.

The pathway is conservative in its data posture. No new biometric collection is mandated beyond what existing digital phenotyping deployments already capture, and no new patient-disclosure burden is imposed beyond the standardized self-report instruments that clinical practice already administers. The architectural value is realized by reorganizing existing inputs under a structural inference layer that satisfies the IEC 62304 lifecycle, ISO 13485 quality system, and ANSI/AAMI EC57 performance expectations through a single artifact. This consolidation is the principal economic and operational argument for the primitive in environments where regulatory burden has historically multiplied parallel instruments without producing a unified clinical state representation.

The architectural posture is consistent across stages: therapeutic monitoring is a continuous structural inference problem, not a periodic instrument-administration problem, and the regulatory record is satisfied through the same primitive that delivers the clinical decision support. Organizations adopting this posture gain a therapeutic monitoring capability that is auditable under HIPAA, validatable under EC57, exchangeable under FHIR R4, and defensible under the FDA AI/ML SaMD and EU MDR regimes, without sustaining the parallel-instrument complexity that procedural compliance has historically required.