Affective State for Elderly Care Companion Agents

1. Regulatory Framework

Elderly-care companion agents operate inside a regulatory perimeter that is more demanding than consumer AI and distinct from clinical decision support. CMS Conditions of Participation for skilled nursing facilities (42 CFR Part 483) impose resident-rights, quality-of-care, and quality-of-life standards that any deployed technology must support rather than degrade; §483.10 resident rights include privacy of personal and clinical communications, which extends to AI-mediated interaction. The Older Americans Act and state Elder Justice frameworks impose additional duties on operators with respect to vulnerable adults, including mandatory reporting of conditions detected during care that suggest neglect, abuse, or significant change in condition.

HIPAA applies to any companion deployed by a covered entity or business associate that records protected health information, including emotional-state data tied to identifiable residents. The FDA has clarified through its 2022 Clinical Decision Support guidance that software which monitors patient state and surfaces alerts to clinicians falls within Section 520(o) Cures Act exclusions only when the clinician can independently review the basis for the alert — a requirement that disqualifies opaque LLM outputs and elevates structured, auditable affective trajectories. The EU AI Act classifies emotion-recognition systems used in workplaces and education as high-risk; deployments in care facilities are not categorically excluded and are likely to fall under the high-risk regime when they materially influence care decisions. State long-term-care licensure regimes (California Title 22, New York 10 NYCRR Part 415, Florida Chapter 400) impose facility-level survey and inspection requirements that any deployed companion must withstand.

Layered on top is a tort-liability environment in which families and state attorneys general scrutinize any technology touching elder care. The combined effect is that an elderly-care companion cannot be a stateless chatbot. It must maintain auditable, governable, structured emotional records that survive a regulatory survey, a §483.10 privacy challenge, an EU AI Act conformity assessment, and a wrongful-death deposition.

2. Architectural Requirement

The architectural requirement is for emotional continuity that is deterministic, governable, and clinically meaningful over the timescales that matter for elder care. Elderly residents, particularly those in assisted living or memory care, depend on consistent relationships for emotional wellbeing. Human caregivers rotate through shifts, introducing discontinuity. Family visits are intermittent. The social isolation that results from inconsistent relationships correlates with cognitive decline, depression, and reduced quality of life — outcomes that CMS surveys, MDS 3.0 assessments, and state ombudsman reviews are explicitly designed to detect.

For the companion to be a genuine care asset rather than a novelty, the architecture must support emotional fields that persist across sessions, evolve over weeks and months, and produce trajectory signals that care staff and clinicians can act on. The fields must be inspectable — a regulator must be able to ask "why did the companion alert" and receive a structured answer. They must be governable — a facility administrator must be able to constrain what the companion is permitted to express, retain, and disclose. They must be tamper-evident — a family member alleging mistreatment must be able to obtain a credentialed history that survives discovery.

For elder care, the emotional trajectory is often more clinically significant than any single interaction. Gradual withdrawal, increasing irritability, declining engagement with previously enjoyed activities, and shifts in sleep-related conversation patterns are early indicators of depression, medication side effects, or cognitive changes. A companion without persistent affective state cannot detect these trajectories, and a companion with persistent affective state but no governance substrate cannot defend its alerts under FDA, AI Act, or tort scrutiny. The architectural requirement is therefore both: persistent state and credentialed governance over that state.

3. Why Procedural Approaches Fail

The procedural approach taken by current companion-AI deployments is to prompt a large language model with a session transcript, optionally appended with retrieved snippets of prior conversation. This works for a single conversational turn and fails as a substrate for elder care. The LLM has no persistent emotional model; its outputs are conditioned on whatever prompt window is currently in scope. A resident who has been gradually withdrawing over two weeks receives the same interaction as a resident who is socially engaged and thriving. The companion cannot detect the trajectory because it has no persistent emotional model — the trajectory exists only in the operator's logs, which are not consulted at interaction time and are not structured for clinical use.

Retrieval-augmented overlays partially address recall but do not produce a governable emotional state. The retrieved fragments are unstructured prose, not named fields with computable update rules. A regulator cannot ask "what is the resident's current contentment baseline" because no such field exists; there is only a corpus of prior text the model may or may not surface. FDA Cures Act exclusion eligibility requires the clinician to "independently review the basis" for a recommendation, which an LLM-generated narrative does not supply. EU AI Act high-risk conformity assessment requires documented data governance, traceability, and human oversight — which an opaque generative pipeline cannot demonstrate.

Wraparound dashboards that score sentiment per session and display a chart over time approximate the trajectory view but inherit two structural defects. First, the score is computed post hoc and does not influence the companion's behavior — the companion remains stateless even when its operator is not. Second, the score is not credentialed; it is an analytics artifact rather than a record that can be admitted as evidence of monitoring duty in a §483.10 dispute. Procedural overlays — operator dashboards, post-hoc reviews, manual care notes — improve outcomes within their operational envelope but cannot supply what the regulatory regime structurally requires: a governed, persistent, inspectable emotional substrate inside the companion itself.

4. The AQ Affective-State Primitive

The Adaptive Query affective-state primitive disclosed under USPTO provisional 64/049,409 supplies a deterministic control substrate of named emotional fields with computable update rules and credentialed governance. An engagement field tracks the resident's social participation level over weeks. A contentment baseline represents the resident's typical emotional state, updated gradually under a published rule. An anxiety field responds to expressed concerns and decays naturally between sessions. A connection field reflects the depth of the relationship between the resident and the companion over time. Each field is a named, typed, persisted variable with a documented update law — not a token in a prompt.

These fields update according to computable rules tuned for elder-care dynamics. Engagement decay is slower than in younger populations because elderly residents may have longer intervals between interactions due to health events. Contentment-baseline shifts are monitored over weeks rather than days, with the rate of change being more significant than the absolute value. The companion's interaction style adapts based on these persistent fields: when engagement is declining, the companion introduces more stimulating conversation topics, references shared memories from previous sessions, or suggests activities; when anxiety is elevated following a health event, the companion provides more reassurance and checks in more frequently. These adaptations occur structurally rather than being prompted per session.

The clinical value of persistent affective state in elder care lies in trajectory detection. When a resident's emotional fields show patterns that correlate with known health concerns, the companion can alert care staff. A sustained decline in engagement combined with increasing anxiety may indicate the onset of depression. A shift from the contentment baseline after a medication change may indicate a side effect. Increasing confusion signals in conversation combined with emotional volatility may suggest cognitive changes requiring assessment. The companion does not make clinical diagnoses. It maintains a computable emotional trajectory and flags deviations from established baselines to care staff, who receive structured emotional data supplementing their clinical observations. The primitive is technology-neutral with respect to update law (any documented rule), composition (additional fields, hierarchical aggregation across residents), and storage. The inventive step under provisional 64/049,409 is the closed governance loop over named, persisted affective fields as a structural condition for emotionally-continuous agents.

5. Compliance Mapping

The affective-state primitive maps directly onto the regulatory regime in Section 1. CMS §483.10 resident-rights privacy is satisfied through field-level governance: emotional data about residents is governed by policy constraints that define who can access trajectory data, how it is used, and what thresholds trigger alerts. Family members may be granted access to summary emotional reports with the resident's consent; care staff receive clinically relevant trajectory alerts; the companion's emotional data is never used for purposes outside the defined care relationship. The named-field architecture makes "purpose limitation" a structural property rather than a policy aspiration.

FDA Cures Act §520(o) clinician-independent-review is satisfied because alerts are generated from inspectable field trajectories with documented update laws — the clinician can see exactly which field crossed which threshold under which rule. EU AI Act high-risk traceability, data-governance, and human-oversight obligations are satisfied through the credentialed lineage record over field updates. HIPAA §164.312(b) audit controls and §164.312(c)(1) integrity are satisfied by the same lineage record. State long-term-care survey requirements and Elder Justice mandatory-reporting duties are supported by structured trajectory exports rather than by ad-hoc operator notes. Valence stabilization ensures the companion maintains psychologically appropriate emotional boundaries — the companion does not develop dependency on the resident's engagement, does not become emotionally volatile when the resident is having a difficult day — which is the structural answer to therapeutic-relationship concerns that consumer chatbots cannot address. Tort posture improves because the credentialed trajectory record is a stronger evidentiary artifact than current care-note systems in any wrongful-death or neglect proceeding.

6. Adoption Pathway

A care operator deploying affective-state companions does not replace its EHR, MDS workflow, or clinical staffing. The primitive integrates as a substrate beneath whatever conversational front-end the operator has selected. The operator defines the field set (engagement, contentment, anxiety, connection, plus operator-specific fields), the update laws (typically authored once with a clinical advisor and versioned thereafter), the governance policy (who reads what, what triggers alerts, what is exported to the EHR), and the credential taxonomy (resident, family-with-consent, nursing staff, attending clinician, facility administrator, regulator).

A staged rollout begins with a single facility and a single field set focused on depression-screening trajectory, where the clinical literature is strongest and the alert-action pathway is clearest. Subsequent stages add medication-change monitoring, cognitive-change trajectory, and family-engagement integration. For educational and family-integration purposes, family members may be granted access to summary emotional reports with the resident's consent, mediated by the same credential taxonomy that governs clinical access. Emotional data about elderly residents requires careful governance, and the credential-mediated field architecture is what makes such governance auditable rather than aspirational.

Commercial framing for the operator is direct: affective-state companions reduce undetected-decline incidents, support MDS 3.0 mood-section accuracy, generate longitudinal data that intermittent-shift staffing structurally cannot, and produce a regulator-defensible posture under CMS, FDA, and EU AI Act regimes. Honest framing — the AQ primitive does not replace caregivers, clinicians, or family. It supplies the persistent, governed, inspectable emotional substrate that turns AI companions from novelty conversation partners into clinical support tools. A caregiver who sees a resident for thirty minutes per shift cannot detect gradual two-week trajectories. A companion that interacts daily and maintains persistent affective state can — and, with the AQ primitive, can do so in a form that survives regulatory and adversarial scrutiny.