A3: Continuous temporal sampling

Operating lifetimes vary by sensor type and follow surprisingly distinct patterns:

Cardiac pacemakers achieve 5-15 years on a single battery using lithium-iodide chemistry chosen specifically for its very low self-discharge rate. Continuous glucose monitors (Dexcom G7, Abbott FreeStyle Libre) maintain clinical-grade accuracy for 7-14 days before sensor replacement is required, with degradation driven by enzyme inactivation and biofouling rather than power loss. Neural interfaces show wider variance: Utah-style microelectrode arrays have demonstrated decade-scale chronic recording in primates but typically lose channels progressively, with usable channel counts often halving within 1-2 years. The Abbott CardioMEMS heart-failure monitor has demonstrated 10+ year reliability in some patients for hemodynamic monitoring.

The failure curve is generally gradual rather than catastrophic — sensors don't suddenly fail; they accumulate fibrous encapsulation and lose accuracy progressively, or individual channels drop out while others continue to function. For AIHS, the operationally relevant question is how sensor accuracy decays over time and at what point the decay matters for the decisions being made from the data. Different clinical decisions tolerate different decay rates; AIHS architecture must include sensor-aging models and confidence weighting that accounts for time-since-implantation.

The foreign body response is the dominant issue across nearly all implantable sensor types. The cascade is well-characterised: neutrophils arrive within hours, macrophages aggregate at the implant surface within days, those macrophages fuse into foreign body giant cells that secrete inflammatory cytokines, fibroblasts are recruited and produce collagen, and within 2-4 weeks a fibrous capsule of 50-200 micrometre thickness isolates the implant from surrounding tissue.

For sensors, this capsule both physically blocks analyte access (slowing diffusion to the sensing element) and changes the local microenvironment (altering pH, oxygen tension, and the very analyte concentrations being measured).

Strategies under investigation include: zwitterionic polymer coatings (polysulfobetaines, polycarboxybetaines) that reduce protein adsorption — the first step of the response — by up to 90%; dexamethasone-eluting coatings that release anti-inflammatory drug continuously and substantially reduce fibrous encapsulation in animal models; softer mechanical interfaces (hydrogels, conductive elastomers) that reduce shear stress on surrounding tissue and provoke less response than rigid silicon or metal; and cell-mimetic surfaces that present molecular signals identifying the implant as "self" to the immune system.

The Langer lab (MIT), Anderson lab (Maryland), Reichert lab (Duke), and several others have made significant progress. The field is moving but no general solution exists yet — the right strategy depends on the implant location, the sensor chemistry, and the required operating lifetime.

Several constraints distinguish medical sensor networks from commercial IoT architecture.

Privacy regulation (HIPAA in the US, GDPR in EU, PIPEDA in Canada, equivalent international frameworks) requires data minimisation, end-to-end encryption, audit trails on every data access, and clear data flows that consumer IoT typically lacks. Cross-jurisdictional data movement is heavily restricted in ways that affect architectural choices.

Latency requirements vary substantially by use case: arrhythmia detection needs sub-second response, sepsis monitoring needs minute-scale, chronic disease monitoring tolerates hours. Edge computation handles latency-sensitive decisions locally; aggregation for population patterns goes upstream over slower paths. Pure-cloud architectures common in consumer IoT cannot meet sub-second medical latency budgets.

Patient data ownership is unsettled — different jurisdictions assign rights differently, and the principle that patients should be able to access, port, and delete their own data conflicts with the operational requirement that clinical decisions made from that data be reconstructable for years for legal and regulatory review.

Clinical accountability requires that any data informing care decisions be reconstructable in detail, meaning permanent storage with versioning, complete audit trails, and traceability from clinical outcome back through every data point and processing step that contributed. The continuous-learning architecture (B5) adds the requirement that models updating from this data must remain auditable for safety.

Together these mean medical sensor networks are substantially more architecturally constrained than commercial IoT and require dedicated infrastructure rather than borrowed consumer-grade systems.