Adaptive Human Interaction Layer

Elizaveta Kutis, MD Founder, 4DOCTORS LLC 

Executive Thesis

Information is not truly delivered when it is sent. It is delivered when it can be received.

Most systems that communicate with humans — electronic health records, patient portals, discharge summaries, conversational AI, service robots — are built around a single design question: what should be said. Content is drafted, approved, and transmitted as though the recipient were a constant: a stable, attentive, cognitively available reader who processes every message the same way, at every hour, regardless of what is happening inside their body or mind at that moment.

This assumption is convenient and almost always false. A person who is exhausted, in pain, anxious, sedated, sleep-deprived, or simply cognitively saturated by the rest of their day does not process a message the way a calm, rested, attentive person does — even when the message is identical. Healthcare in particular produces enormous quantities of correct information that nonetheless fails to become action, not because the medicine was wrong, but because the communication was not matched to the person receiving it.

This article proposes a research direction we call the Adaptive Human Interaction Layer — a modular translation layer, sitting between systems that estimate human physiological and contextual state and systems that communicate with humans, whose function is to answer a question current technology rarely asks:

Given what this person can plausibly process right now, how should this information be delivered?

The underlying content does not change. The delivery does. This distinction — between adapting what is true and adapting how truth reaches a person — is the foundation of everything that follows, and it is the reason this work is framed as engineering and communication science rather than persuasion.

The Missing Layer in Human-Machine Communication

Two research programs have matured over the past three decades, largely in parallel.

The first estimates human states from signals. Affective computing, a field formally proposed by Rosalind Picard in 1997, defined computation "that relates to, arises from, or deliberately influences emotions," and established the broader project of inferring affective and physiological states from sensor data. Physiological computing extended this tradition by treating signals recorded directly from the body — cardiac, electrodermal, respiratory — as inputs to interactive systems. Digital phenotyping, formalized by Onnela and colleagues, describes the "moment-by-moment quantification of the individual-level human phenotype in situ" using data from personal devices, and has been applied to conditions ranging from depression risk to post-surgical recovery.

The second adapts systems to context. Context-aware computing, defined by Dey and Abowd as systems that use "any information that can be used to characterize the situation of an entity" to provide relevant services, gave rise to adaptive user interfaces that change content, layout, or behavior based on location, task, or device. Cognitive load theory, formulated by Sweller, established that working memory has a bounded capacity, and that instructional design should minimize extraneous load so that the mind's limited resources can be spent on genuinely useful processing.

What is largely missing is the layer connecting the two: a system that takes an estimate of a person's current physiological and cognitive availability and uses it to systematically restructure how information is delivered to that person, across whatever interface is in use — a screen, a voice assistant, a robot, a wearable notification.

We are not proposing that this idea has never been touched. Pieces of it exist across affective computing, human-robot interaction, personalized health communication, and adaptive interface research, cited throughout this article. What we propose is a dedicated, general-purpose translation layer — middleware, not a single application — whose sole responsibility is this transformation, and whose value proposition is explicit: recognition is only the first half of intelligent interaction. Adaptation is the second.

From Physiological Interpretation to Adaptation

Any system built on this idea depends on a prior, separately validated stage. Before delivery can be adapted, the system must have some estimate of the person's current state — Stage 1 in a two-stage architecture:

Stage 1 — Interpretation. Physiological and contextual signals (heart rate variability and other ECG-derived features, heart rate, respiration, sleep and recovery metrics, activity, electrodermal activity where available, voice characteristics, behavioral signals, self-report, and situational context) are processed into an estimate of the person's current state.

Stage 2 — Adaptation. That estimate is used to shape how information is subsequently communicated.

This article is concerned with Stage 2. But Stage 2 is only as trustworthy as Stage 1, and the two must not be blurred. Heart rate variability, for example, is a well-studied marker of autonomic nervous system activity, and a growing body of wearable-sensor research combines it with respiration, electrodermal activity, and accelerometry to improve robustness of state estimation. This is established, active science. It is a different — and much stronger — claim than saying HRV alone reveals that a specific person is "angry" or "burned out." The architecture we describe treats physiological signals as one input among several into a probabilistic, multimodal estimate of state, never as a standalone emotional diagnosis.

State Is Not Identity

The architecture separates three constructs that are often collapsed into one in casual discussion, and should not be:

Physiological state — what appears to be happening in the organism right now: elevated sympathetic activity, fatigue, autonomic recovery, physical strain.

Emotional or affective context — how the person may currently be experiencing the interaction: uncertain, reassured, overwhelmed, calm. This is inferred with lower confidence than physiological state and should be treated as a hypothesis, not a label.

Dynamic Interaction Profile — how information can most effectively be structured for this person, encompassing both a momentary component (what can this person process right now) and a more stable component (how does this person generally prefer to receive information).

These three layers are not equivalent, they are not equally certain, and a well-designed system must represent that difference honestly rather than presenting a single confident number to the person or the clinician on the other end. State is a signal, not an identity, and it is expected to change from one hour to the next.

How the Adaptive Human Interaction Layer Works

The transformation the architecture performs can be represented as a pipeline:

PHYSIOLOGICAL + CONTEXTUAL INPUTS
        ↓
STATE INTERPRETATION
        ↓
ADAPTIVE HUMAN INTERACTION LAYER
        ↓
COMMUNICATION TRANSFORMATION
        ↓
CLINICAL / AI / ROBOTIC / DIGITAL INTERFACE
        ↓
HUMAN COMPREHENSION + ACTION
        ↓
FEEDBACK

The layer sits between state estimation and whatever is doing the communicating. It does not originate clinical content and it does not alter facts. What it can change, for a given piece of information, includes: message length, vocabulary complexity, the number of steps presented at once, sequencing, tone, the amount of explanation offered, repetition, visual hierarchy, urgency framing, use of examples, choice architecture, timing (immediate versus deferred delivery), whether confirmation of understanding is requested, and whether the person is shown one action or several options at once.

Consider a physician who wants a patient to (1) obtain lab work, (2) schedule imaging, (3) adjust a medication schedule, and (4) return in two weeks. The medical instruction is identical for every patient who receives it. But the way it should arrive plausibly differs by state.

A patient who is calm, rested, and cognitively available might reasonably be given the full explanation at once: rationale, alternatives, scheduling logistics, and educational background, presented together because that patient can hold and use all of it.

A patient under high physiological strain, or with low apparent cognitive availability, may be better served by a narrower first message — "Today, do only this: schedule your blood test" — with the remaining steps deferred and delivered in sequence once the first is confirmed.

Neither patient receives different medical information. Both receive the same four instructions eventually. What differs is the sequencing and density of delivery — and the purpose is not to manipulate either patient toward a particular choice, but to reduce unnecessary cognitive burden while preserving the accuracy, autonomy, and completeness of what each of them is ultimately told.

Healthcare: From Correct Information to Executable Information

Healthcare is an unusually strong proving ground for this idea because the gap between correct information and acted-upon information is large, well documented, and expensive.

Roughly one in five new prescriptions in the United States is never filled, and of those that are filled, roughly half are not taken as prescribed with respect to timing, dosage, frequency, or duration, according to the CDC's 2017 Grand Rounds synthesis, which itself draws these figures from earlier adherence research rather than presenting new CDC data. That same synthesis reports direct nonadherence-related healthcare costs in the range of $100–300 billion annually in the U.S. — a range that recurs across analyses spanning more than a decade and traces back to earlier NEHI and pharmacoeconomic estimates. This is a distinct and narrower claim than the $528.4 billion figure reported by Watanabe, McInnis, and Hirsch (2018), which measures a broader category — the cost of nonoptimized medication therapy overall, including treatment failures and new medical problems arising from suboptimal prescribing, not nonadherence alone. The two figures should not be conflated: one estimates the cost of patients not taking medication as prescribed; the other estimates the cost of medication therapy not being optimized across several distinct causes, of which nonadherence is only one.

Health literacy compounds the problem. AHRQ's synthesis of the National Assessment of Adult Literacy found that roughly 36 percent of U.S. adults have basic or below-basic health literacy, and separate AHRQ materials estimate that as many as 88 percent of adults lack the health literacy skills needed to manage the full demands of the current healthcare system. Limited health literacy has been associated with higher hospitalization and emergency-care use, poorer ability to interpret medication labels and health messages, and, among older adults, higher mortality.

Discharge communication makes the mechanism concrete. A study of hospitalized children found comprehension error rates as high as 78 percent for return-precaution instructions, and demonstrated that discharge-plan complexity and low caregiver health literacy each independently predicted comprehension and adherence errors — with the two compounding when combined. A separate analysis found that a majority of patients had functional reading skills below the reading level of their own discharge paperwork. None of this reflects incorrect medicine. It reflects a mismatch between the density and structure of the message and the processing capacity of the person receiving it at the moment they received it.

Framed against this backdrop, an adaptive communication layer is not a comfort feature. It is a plausible intervention against a well-quantified, expensive, and currently under-addressed failure mode in healthcare delivery — one that sits downstream of correct clinical decision-making and upstream of the patient behavior that decision-making is meant to produce.

We want to be precise about what has and has not been shown. It has not been demonstrated that an adaptive communication layer of the kind described here improves comprehension, recall, adherence, or any other outcome in a validated clinical trial. What the literature above establishes is the size and mechanism of the problem this architecture is designed to address. Comprehension, recall, task completion, adherence, engagement, reduction of cognitive burden, communication errors, missed appointments, order completion, trust, and satisfaction are the measurable hypotheses this architecture generates — not results it has already produced.

The Dynamic Interaction Profile

Momentary state answers one question: what can this person reasonably process right now? A second, slower-moving layer answers a related but distinct question: how does this person generally process information most effectively?

It is tempting to describe this second layer only as a matter of how much information to give — more for an available person, less for a depleted one. That framing is incomplete. What the architecture proposes is broader: the form of the interaction itself is a variable the system can adjust, not only its length or density. For illustration, an adaptive system might select among communication modes such as:

Direct — the action stated first, with no surrounding explanation unless requested. "Your lab order is ready. Schedule it here."

Explanatory — the rationale precedes the action, because some people will not act on an instruction whose mechanism they do not understand. "Your physician ordered this test because it will help evaluate X. Here is what will happen and why it matters."

Structured — the same content decomposed into discrete, numbered steps, favoring visual hierarchy over narrative flow. "Step 1 — Schedule lab. Step 2 — Complete imaging. Step 3 — Follow up."

Conversational — a calmer, reassurance-forward register, suited to a person who needs to feel accompanied through a decision rather than instructed.

This set is a proposed, illustrative taxonomy — one candidate way of organizing the design space — not a set of established scientific categories, and we do not claim any published literature has isolated exactly these four modes as a validated classification. The substantive claim is narrower and, we think, more defensible: that the mode of interaction, not just its volume, is itself a variable the system can adjust, and that the right mode for a given person is not fixed. The same individual may need a Direct mode on a day of high physiological strain and an Explanatory mode three days later, once recovered, when they are ready to understand the reasoning behind a decision they accepted quickly the first time.

We refer to the combination of a person's observed interaction-mode preferences, refined over repeated exchanges, as a Dynamic Interaction Profile. We deliberately avoid framing this in terms of fixed personality typologies unsupported by evidence, and we avoid any term that would imply a stable, biologically or clinically validated trait — the profile is a learned, revisable pattern the system maintains about how a given person tends to engage, not a diagnosis or a fixed classification of the person themselves. It is an open research problem at the intersection of communication preference, health literacy, cognitive load, and behavioral context, built and refined only through repeated interaction, never assigned from a one-time questionnaire. Momentary state tells the system how much a person can take in right now; the Dynamic Interaction Profile tells it which mode is likely to land — and the two combined are expected to produce a more useful adaptation than either alone.

"Adapt to Me": A Universal Interaction Primitive

A simple, memorable interface pattern makes the idea concrete, and it is worth making concrete at length, because it is arguably the clearest single demonstration of what this architecture does — one that requires no mention of HRV, sensors, or any physiological detail to be understood.

Imagine an email, a portal message, or a chat interface with a button labeled Adapt to Me. The factual content behind the button never changes. What changes is its presentation. A single underlying instruction — the same clinical order, the same policy update, the same appointment reminder — can be rendered in whichever mode fits the recipient and the moment.

Take a routine, formally worded patient message as the source content:

"Dear Patient, please be advised that pursuant to the recommendations discussed during your recent appointment, laboratory testing has been ordered to further evaluate your current condition. Please contact our office at your earliest convenience to arrange scheduling."

Pressed by one recipient, Adapt to Me could render this as:

You have one thing to do today: Book your blood test. [Schedule]

Pressed by a different recipient — someone who engages better once they understand the mechanism — the same underlying order becomes:

Why your doctor ordered this test This test measures X and helps determine Y. After the result, the next decision will be Z. [See my order]

Both renderings originate from the same clinical order. Neither adds, omits, or softens a fact. What differs is which of the interaction modes described above the system reaches for, and how much is surfaced in the first screen versus made available on request. This is, deliberately, the least technical possible illustration of the entire architecture — it sells the idea without requiring the viewer to understand or trust any sensor input at all, which is precisely why it is a useful entry point for an audience evaluating the concept rather than the underlying signal science.

The same primitive generalizes far beyond a single inbox: to messaging apps, patient portals, discharge instructions, telemedicine, AI assistants, education platforms, workplace communication, customer service, emergency and public-safety communication, robotics, vehicle interfaces, wearables, elder care, rehabilitation, and safety-critical domains such as aviation and first-response work. In every one of these settings, the system is currently asking only "what information should I give this person?" We propose that it should also ask: what is the most appropriate way to deliver this information to this human, in this state, at this moment?

From Screens to Robots

Today, most AI systems communicate with people through a screen. That will not remain true. Clinical robots, elder-care robots, rehabilitation systems, service robots, autonomous vehicles, and home and workplace robots are increasingly physically present participants in human environments, and the research literature on human-robot interaction already treats this seriously.

Affect-adaptive human-robot interaction is an active research area in which robots "perceive, model, and dynamically adapt [their] behavior to the affective state of the human user," using multimodal cues, closed-loop adaptation, and personalized feedback. Work on socially assistive robotics for older adults has shown that robots equipped with emotion-aware dialogue produced measurably different engagement than scripted, non-adaptive versions of the same robot in the same setting. Frameworks grounded in models such as the ICF (International Classification of Functioning, Disability and Health) have been proposed specifically to let assistive robots personalize both content and behavior to the evolving condition of the person in front of them, particularly frail older adults whose needs and states change over time. Studies of older adults' expectations of care robots consistently identify slower, clearer speech, reduced technical vocabulary, and an emotionally supportive but non-patronizing tone as prerequisites for acceptance — precisely the dimensions this architecture is designed to modulate.

A future clinical or care robot that detects — through whatever sensing is available to it — that a patient is under high physiological strain should not respond by simply repeating itself more loudly or more often. It could instead shorten its sentences, slow its speech rate, reduce the number of questions or choices offered at once, insert pauses, request confirmation before proceeding, and escalate to a human clinician when appropriate. The Adaptive Human Interaction Layer is best understood, in this context, as middleware sitting between sensors and human-state estimation on one side, and AI, robot, or digital interface output on the other — infrastructure, not a single application, and not "another chatbot."

Platform Architecture and Commercial Applicability

The central architectural claim of this section is easy to state and easy to underestimate: the Adaptive Human Interaction Layer does not need to own the application it improves.

It is not an EHR, a robot, an AI assistant, an email client, or a wearable. It is the decision layer that sits between an estimate of human state and whatever interface already exists:

HUMAN STATE  →  4DOCTORS ADAPTATION ENGINE  →  EXISTING APPLICATION

Read literally, this means the layer does not require replacing an EHR to improve how that EHR's portal messages are delivered. It does not require manufacturing a robot to change how an existing clinical or care robot paces and sequences its speech. It does not require building a competing AI assistant to change how an existing one structures its answers. It does not require building an email client to power the "Adapt to Me" button described above. It does not require producing a wearable to consume the physiological signal a wearable already provides. In every one of these cases, the incumbent system keeps doing what it already does — originating content, running the underlying application logic, owning the relationship with its user — and the adaptation engine supplies one specific decision: how should this piece of content be shaped for this person, right now, given what is known about their state and their Dynamic Interaction Profile.

This is a structurally different commercial position than building another point solution. A point solution competes for a category an incumbent already occupies. A decision layer of this kind is additive to every incumbent in a category at once — an EHR vendor, a robotics company, a wearable maker, a conversational-AI platform, and an emergency-communication system can each integrate it without giving up ownership of their own application or user relationship. That is what makes it plausible as infrastructure rather than as a single product: the value is concentrated in the decision the layer makes, not in the surface the person ultimately sees or clicks.

Consistent with that framing, business models worth exploring include an API or SDK exposing the adaptation decision to any integrating application, licensing of validated adaptation models to platforms that already own an interface, and domain-specific adaptive-communication engines tuned to a single vertical such as discharge communication or elder care. We deliberately do not attach a market-size figure to this vision; speculative total-addressable-market estimates would add nothing to its credibility. What can be said with evidence is that the underlying problems this layer targets are large and well documented — hundreds of billions of dollars in nonadherence-related costs annually in the U.S. alone, tens of millions of adults with limited health literacy, and comprehension-error rates in discharge communication reaching into the double digits even for the return-precaution instructions patients are specifically warned to remember (see References). A layer that meaningfully improves the rate at which correct information becomes executed action is addressing a problem of that scale — not inventing a market, but attaching engineering to one that clearly already exists, and doing so without requiring any single incumbent to be displaced first.

Scientific Foundation and Related Fields

This proposal draws on, and should be read alongside, several established and active research areas: affective computing and physiological computing (state inference from physiological and behavioral signals); context-aware computing and adaptive user interfaces (systems that change based on situational information); cognitive load theory (the bounded capacity of working memory and the design implications that follow from it); personalized health communication and health literacy research (matching message complexity to reader capability); digital phenotyping (continuous, passive quantification of behavioral and physiological signals from personal devices); multimodal sensing for stress and affect detection (combining heart rate variability with respiration, electrodermal activity, and accelerometry for more robust state estimates); and human-robot interaction, particularly socially assistive robotics and affect-adaptive HRI.

We are not claiming that adaptive communication is a novel idea in the abstract — each of these fields has touched pieces of it. What we believe is distinct is the proposal of a dedicated, modular translation layer that sits explicitly between physiological-state interpretation and the adaptation of communication, designed to generalize across interfaces and domains rather than being built into any single application. Put simply: we are not primarily interested in recognizing the human. We are interested in what happens to communication after recognition — and building that transformation as reusable infrastructure rather than as a one-off feature of any single product.

What Must Be Validated

For this architecture to be trustworthy rather than merely appealing, its two stages must be validated separately and in sequence, because an adaptation built on an unreliable state estimate is not a shortcut — it is a source of new errors.

1. Validate the underlying physiological and contextual measurements — establish that the signals used (HRV, respiration, and so on) reliably track the physiological constructs they claim to track, in the populations and conditions where the system will be used.
2. Determine which measured states meaningfully affect information processing — not every physiological fluctuation changes what a person can comprehend; this must be established empirically, not assumed.
3. Define adaptation rules or models — translate state estimates into concrete, testable delivery decisions.
4. Compare standard versus adaptive communication — controlled comparison against the status quo, not against a strawman.
5. Measure comprehension, recall, task completion, adherence, cognitive burden, and user preference as primary outcomes.
6. Test generalizability across individuals, literacy levels, cultural contexts, and populations — an adaptation tuned to one group may not transfer to another.
7. Validate in healthcare settings specifically, given the stakes and the existing evidence base described above.
8. Extend to other human-machine interaction environments only after healthcare validation, not in parallel with it.

This sequencing matters. The claim we are prepared to defend at this stage is that physiological interpretation must be established as reliable before any downstream communication adaptation built on it can be trusted — not that the downstream adaptation already works.

Ethical Boundaries

Adaptive communication changes how something is said. It must never become a covert way of changing what is effectively communicated, or of exploiting a person's state to obtain a particular decision from them. We treat the following as non-negotiable design constraints, not aspirational values:

Informed consent for any use of physiological or contextual data to shape communication.

Privacy of physiological data, with clear boundaries on collection, storage, and use.

User control, including a straightforward ability to disable adaptation entirely and receive information in a standard, unadapted form.

Transparency about the fact that adaptation is occurring, and, where relevant, about what triggered a given adaptation.

Attention to algorithmic bias, particularly given that health literacy and access to fluent smartphone or wearable use are unevenly distributed across populations.

Accessibility as a first-class design requirement, not an afterthought.

Preservation of autonomy — adaptation should support a person's ability to understand and choose, never substitute for it.

Preservation of factual content — the clinical or informational substance delivered must remain complete and accurate regardless of how delivery is adapted.

Avoidance of exploiting emotional vulnerability — a state estimate indicating distress must never be used to increase persuasive pressure.

Appropriate clinical oversight wherever the content being adapted is medical.

The organizing principle for all of this is simple to state and demanding to implement: adapt the delivery, not the truth. The system may change how information is presented. It must not change what is true, or make information more persuasive by any means other than making it more comprehensible.

Research Roadmap

The path from concept to validated system runs through the eight stages above, applied first narrowly and then broadly:

Begin with a single, well-scoped clinical communication task (for example, post-visit instruction delivery) where ground truth for comprehension and adherence can be measured directly.

Validate the physiological or contextual signal being used before layering any adaptation logic on top of it.

Run controlled comparisons of adapted versus standard delivery on that single task, using comprehension, recall, and task completion as primary outcomes.

Only after a positive, replicated result in that narrow setting, extend to additional tasks, populations, and eventually additional interface types — digital, conversational, and, later, embodied and robotic.

Each stage is intended to be falsifiable: a negative result at any stage — no measurable difference between adapted and standard delivery, or a state estimate that does not meaningfully predict processing capacity — should stop further extension until the underlying assumption is revisited.

Future Vision

Personalized medicine has, so far, largely meant personalizing treatment: the dose, the drug, the protocol matched to the individual. It has not, in general, meant personalizing communication — the message that tells a person what to do and why. As AI systems take on a larger share of the work of communicating with people — through screens today, through embodied and robotic systems increasingly tomorrow — the question of how much a given person can process at a given moment becomes a design question those systems cannot keep ignoring.

The next generation of intelligent systems may need to understand not only what to say, but how much the human in front of them can process right now. A clinically correct instruction that cannot be processed is still a communication failure. The Adaptive Human Interaction Layer is our proposed name for the research direction, and eventually the infrastructure, aimed at closing that gap — deciding not only what information a person needs, but how that information should reach that particular human, in that particular state, at that particular moment.

References

1. Watanabe, J. H., McInnis, T., & Hirsch, J. D. (2018). Cost of Prescription Drug–Related Morbidity and Mortality. Annals of Pharmacotherapy, 52(9), 829–837. https://doi.org/10.1177/1060028018765159 — Source for the $528.4 billion (2016 USD; range $495.3–672.7B) estimate of the annual cost of nonoptimized medication therapy, equivalent to 16% of 2016 U.S. healthcare expenditure.

2. CDC. (2017). CDC Grand Rounds: Improving Medication Adherence for Chronic Disease Management — Innovations and Opportunities. MMWR Morbidity and Mortality Weekly Report, 66(45), 1248–1251. https://www.cdc.gov/mmwr/volumes/66/wr/mm6645a2.htm — Source for the "roughly one in five new prescriptions never filled" and "roughly 50% not taken as prescribed" figures, and the $100–300 billion range for direct nonadherence-related healthcare costs.

3. Kutner, M., Greenberg, E., Jin, Y., & Paulsen, C. (2006). The Health Literacy of America's Adults: Results From the 2003 National Assessment of Adult Literacy (NCES 2006-483). U.S. Department of Education, National Center for Education Statistics. https://nces.ed.gov/pubs2006/2006483.pdf — Source for the 36% of U.S. adults with basic or below-basic health literacy.

4. Brega, A. G., Barnard, J., Mabachi, N. M., Weiss, B. D., DeWalt, D. A., Brach, C., Cifuentes, M., Albright, K., & West, D. R. (2015). AHRQ Health Literacy Universal Precautions Toolkit, 2nd ed. (AHRQ Publication No. 15-0023-EF). Agency for Healthcare Research and Quality. https://www.ahrq.gov/sites/default/files/publications/files/healthlittoolkit2.pdf — Source for the estimate that 88% of U.S. adults lack the health literacy skills needed to manage the full demands of the current healthcare system.

5. Glick, A. F., Farkas, J. S., Mendelsohn, A. L., Fierman, A. H., Tomopoulos, S., Rosenberg, R. E., Dreyer, B. P., Melgar, J., Varriano, J., & Yin, H. S. (2019). Discharge Instruction Comprehension and Adherence Errors: Interrelationship Between Plan Complexity and Parent Health Literacy. Journal of Pediatrics, 214, 193–200.e3. https://doi.org/10.1016/j.jpeds.2019.04.052 — Source for the discharge-instruction comprehension error rates (50% medication side effects, 34% medication dose adherence, 78% return-precaution instructions) and the finding that plan complexity and caregiver health literacy independently predicted comprehension and adherence errors.

6. Picard, R. W. (1997). Affective Computing. MIT Press. — Source for the founding definition of affective computing as "computing that relates to, arises from, or deliberately influences emotions."

7. Dey, A. K. (2001). Understanding and Using Context. Personal and Ubiquitous Computing, 5(1), 4–7. — Source for the definition of context-aware computing as systems using "any information that can be used to characterize the situation of an entity."

8. Sweller, J. (1988). Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science, 12(2), 257–285. — Source for the foundational claim that working memory has a bounded capacity, underlying cognitive load theory.

9. Onnela, J.-P., & Rauch, S. L. (2016). Harnessing Smartphone-Based Digital Phenotyping to Enhance Behavioral and Mental Health Care. Neuropsychopharmacology, 41(7), 1691–1696. — Primary source for the definition of digital phenotyping as the "moment-by-moment quantification of the individual-level human phenotype in situ."

10. Rossi, S., et al. A dichotomic approach to adaptive interaction for socially assistive robots. PMC9670074. https://pmc.ncbi.nlm.nih.gov/articles/PMC9670074/
11. Artificial Emotional Intelligence in Socially Assistive Robots for Older Adults: A Pilot Study. PMC10569155. https://pmc.ncbi.nlm.nih.gov/articles/PMC10569155/
12. Getson, C., & Nejat, G. (2024). Investigating Persuasive Socially Assistive Robot Behavior Strategies for Sustained Engagement in Long-Term Care. arXiv:2408.14322.
13. Acceptability and Usability of a Socially Assistive Robot Integrated With a Large Language Model in a Geriatric Care Institution. PMC12357123. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12357123/ — Source for older adults' documented preference for slower speech, reduced technical vocabulary, and an emotionally supportive tone from care robots.

14. Applications of Heart Rate Variability Metrics in Wearable Sensor Technologies: A Comprehensive Review. Electronics (MDPI). https://www.mdpi.com/2079-9292/15/8/1707
15. State-of-the-Art of Stress Prediction from Heart Rate Variability Using Artificial Intelligence. Cognitive Computation. Springer. https://link.springer.com/article/10.1007/s12559-023-10200-0