The Internet of Medical Things is what happens when the Internet of Things grows up inside a hospital. Devices, sensors, wearables, and imaging hardware stop being islands and start behaving as one connected fabric — feeding patient data into shared infrastructure that clinicians can act on. The interesting question is not whether this network exists. It does. The question is which slices of it are mature enough to rely on, and which are still aspirational. What is the Internet of Medical Things? IoMT is the application of IoT principles — networked devices, sensors, and shared data planes — to medical equipment. A bedside monitor, a continuous glucose sensor, a portable ultrasound probe, and a smartwatch heart-rate sensor all become nodes that emit data the rest of the system can read. The earliest IoT proof of concept dates to 1982, when a soft drink vending machine at Carnegie Mellon was modified to report its inventory and the temperature of its stock. The medical version of that idea took longer to arrive, partly because regulators and partly because the data is harder. Today IoMT spans inpatient telemetry, ambulatory monitoring, consumer wearables, and emerging device classes like ingestible sensors. Figure 1 — The 1982 vending machine widely cited as the first networked-device proof of concept that IoT later formalised. In our experience supporting healthcare and life-sciences teams, the practical definition of IoMT is narrower than the marketing definition. It is the subset of connected medical devices whose data actually reaches a clinical decision-maker in a usable form. Everything else is a sensor that happens to have a radio. Inpatient monitoring: where IoMT already pays for itself The strongest IoMT use case today is continuous inpatient monitoring. Vitals — heart rate, respiratory rate, blood pressure, oxygen saturation, temperature — stream from bedside devices to a hospital server, where clinicians read them on phones, tablets, and workstations. The clinically meaningful pattern is not that the data exists. It is that edge computing in healthcare has matured enough to do the early filtering at the bedside, so the network only carries the events that matter. This shifts the failure mode of remote patient monitoring. Ten years ago the bottleneck was bandwidth and storage. Today the bottleneck is signal-to-noise: every alarm a clinician acknowledges that turned out to be irrelevant erodes attention for the alarm that is not. Edge-side inference on the monitor itself, rather than centralised batch analysis, is the architectural choice that makes the difference. The same fabric supports video consultations. A web camera on the ward, paired with a local network and access controls, lets specialists join a case without flying in. For follow-up consultations, where a patient might otherwise travel internationally to see a particular surgeon, the consultation can happen from home and the hospital keeps the recording and structured notes alongside the rest of the chart. Figure 2 — A clinician reviewing remote telemetry on a portable device. The screen is the visible part; the edge inference layer that filters the signal is not. Computer Vision (CV) sits naturally on top of this video layer. The clinically interesting use is not aesthetic — checking whether a patient looks tired — but structured: pose estimation for fall risk on geriatric wards, gait analysis post-orthopaedic surgery, and behavioural cues that correlate with delirium onset. None of these replace clinical judgement. They produce a second stream of structured observation that a single clinician on rounds could not maintain continuously. Language is the other gap. The communication mismatch between a clinician and a patient — particularly when the encounter is short, remote, or across a language barrier — is a documented contributor to readmission. Natural language processing now handles real-time translation and structured summarisation of the consultation transcript well enough to be useful in the loop, provided the system makes its outputs auditable. We treat Generative AI in this context as a translation and summarisation layer, not as a clinical reasoning layer. What does IoMT look like for chronic conditions? Chronic disease is where IoMT shifts from an inpatient nicety to an outpatient necessity. Cancer, heart disease, and diabetes are conditions where the relevant signal arrives over months, not minutes. The patient is not in a hospital. The device has to do more on its own. Continuous glucose monitoring is the cleanest example. A sensor sits under the skin, samples interstitial fluid at a fixed cadence, and transmits readings wirelessly to a paired device. The clinician sees the trend, not the spot reading. This changes how dose adjustments are decided. Portable cardiac monitors do the same for arrhythmia detection — they hold a rolling buffer, flag events that meet a learned signature, and surface only those to the reviewing physician. Figure 3 — A continuous glucose monitor. The sensor under the skin is the small part; the data pipeline downstream is what makes it clinically useful. Portable imaging is the next frontier. Handheld ultrasound probes priced for general practice exist now, and the imaging quality is good enough that the bottleneck is interpretation. This is where CV becomes load-bearing. Annotating medical images for abnormalities is tedious work that gets worse with fatigue, and human reviewers make more errors on the fortieth scan of a shift than the first. A trained CV model does not fatigue. It does, however, fail in ways humans do not — which is why the deployment pattern that works is assist-and-review rather than autonomous read. Quick-reference: which IoMT applications are operationally mature? Application Maturity What clinicians actually rely on Inpatient vitals telemetry Production Continuous streams, edge-filtered alerts, EHR integration Continuous glucose monitoring Production Trend data informing insulin dosing decisions Remote cardiac event monitoring Production Event-flagged ECG snippets, not raw streams Tele-consultation video Production Routine for follow-up and specialist input CV-assisted medical image review Validated assist Triage and second-read, not autonomous diagnosis Consumer-grade smartwatch vitals Indicative only Trends, not clinical decisions Ingestible sensor pills Early deployment Adherence verification for specific indications AR/VR surgical training Production for training Skill rehearsal, not patient-facing Remote-operated surgery at scale Research Not a routine clinical pathway The line that matters is the third column. A device is part of IoMT in a meaningful sense when a clinician changes a decision based on its output. Everything else is data exhaust. Where do wearables fit in IoMT? Smartwatches are the most visible IoMT category and the most frequently misread. Heart-rate, activity, and sleep tracking are not FDA-cleared as diagnostic measurements, and the accuracy of consumer-grade optical sensors degrades with skin tone, motion, and fit. They are useful as trend indicators and as triggers for medical evaluation, not as replacements for medical-grade devices. The features that have crossed the threshold into clinical relevance are narrow: single-lead ECG capture on certain models, which has FDA clearance for atrial-fibrillation flagging; fall detection with SOS escalation, which has demonstrated value for elderly users living alone; and post-event timeline reconstruction, where the watch happens to be the only device that recorded what was happening when symptoms started. Figure 4 — A smartwatch displaying a real-time heart-rate trace. Useful for trends; not a substitute for a 12-lead ECG. The architecture worth noticing is that smartwatches do meaningful inference on-device. Activity classification, heart-rate variability decomposition, and sleep-stage estimation run on the watch itself, with only the summaries syncing to a phone. This is edge computing at jewellery scale, and it is what makes the always-on use case battery-feasible. The next layer: ingestibles, AR, VR The frontier of IoMT is where new device classes change what the network can sense. Smart ingestible pills — drugs with an embedded ingestible sensor — verify that a medication was taken and absorbed, and transmit that confirmation before the sensor itself is broken down. This is a narrow but high-value capability for indications where adherence determines outcome: tuberculosis treatment, schizophrenia maintenance, complex post-transplant regimens. Figure 5 — A concept rendering of an ingestible sensor pill. The clinical value is adherence verification, not telemetry. Augmented and virtual reality occupy a different slot. Their highest-leverage IoMT application is training, not patient care. A 2009 study cited in The Guardian found that nearly 12% of medical students experienced fainting or near-fainting during operating theatre exposure. Putting the first exposure in VR is not a gimmick — it is the same logic as flight simulators for pilots. Surgical rehearsal in VR has matured to the point where it is being used at named institutions for procedure-specific training, and the underlying GPU and AR/VR infrastructure that makes it interactive is the same fabric powering medical visualisation more broadly. Figure 6 — VR surgical training in use at St George's University Hospital. The clinical value is training without patient risk, not remote surgery. Remote-operated surgery is the application most often discussed and the least mature. The networking and edge requirements are severe — sub-50-millisecond round-trip latency, jitter floors that consumer infrastructure cannot meet, fail-safe handoff that does not yet have a credible standard. We treat it as a 10-year horizon application, not a 12-month one. How should a healthcare team approach IoMT? The methodology that works is the opposite of the marketing pitch. Start with the clinical workflow where data is already being collected and where a single missed signal has high cost. That is usually inpatient telemetry, chronic-disease follow-up, or imaging triage. Connect the existing devices into a single readable plane before adding new device classes. Build the edge inference layer that suppresses noise before that noise reaches the clinician. Only then expand to consumer devices and emerging form factors. The failure mode we see most often is the inverse: a team buys a new device class because it is exciting, plugs it into an inadequate data plane, and discovers that the bottleneck was never the sensor. It was always the integration. FAQ Which AI use cases in pharmaceutical manufacturing are already proven in production today? The proven slots are visual inspection (computer vision replacing manual particulate and cosmetic checks on the line), predictive maintenance on critical equipment such as compression presses and fillers, deviation triage that classifies and prioritises quality events, and process control around continuous manufacturing lines. These are operational use cases, not pilots — they have measurable throughput and quality outcomes attached. Where on the manufacturing line does AI deliver measurable ROI — inspection, deviation triage, predictive maintenance, batch release? Inspection and predictive maintenance produce the fastest payback because the failure mode they prevent — rejected batches and unplanned downtime — is directly quantifiable. Deviation triage compounds over time as the model learns the site’s recurring patterns. Batch release acceleration is real but constrained by validation work; it is a longer ROI horizon, not a shorter one. What separates the proven use cases from the still-experimental ones? The proven use cases share three properties: the input data is already being collected for another reason, the decision the model supports has a human in the loop, and the cost of a false negative is bounded. Experimental use cases typically fail at least one of these — they require new instrumentation, propose autonomous action, or operate where the worst case is unbounded. How are existing pharma AI deployments structured to satisfy GMP and GxP requirements? The pattern we see is to deploy the AI as a decision-support layer with deterministic audit trails, keep the human qualified-person sign-off in the workflow, validate the model as a piece of computer system equipment under GAMP-5, and treat retraining as a change-control event. The model does not own the decision; it owns the evidence the decision is made against. Which use cases are pharma companies abandoning, and why? The ones being quietly dropped are end-to-end autonomous batch release, generative chemistry without wet-lab validation in the loop, and free-text deviation auto-resolution. The common thread is that they all assume a quality of input data that real plants do not produce, and they all expose the manufacturer to regulatory exposure the model cannot carry. What does a credible AI roadmap for a pharma plant look like over the next 12 months? Quarter one is data plane work — getting line data into a shared, queryable form. Quarter two is the first inspection or predictive-maintenance deployment on a single line, with a clear baseline. Quarter three is extending the working model to a second line and starting deviation-triage scoping. Quarter four is validation and computer-system-validation work for whichever model is moving toward GxP-touching use. Anything more ambitious tends to slip, because the underlying data and validation work is what actually paces the programme. Where this leaves IoMT IoMT is not one thing. It is a stack of capabilities at different maturity levels, sitting on a shared data and edge infrastructure. The hospitals and life-sciences teams getting value from it today are the ones treating it as infrastructure to be earned, not a product to be bought. They invest in the edge layer, they integrate the devices they already have before adding new ones, and they put the human clinician at the centre of the decision loop rather than at the edge of it. At TechnoLynx we work with healthcare and life-sciences teams on exactly this layer: the integration, edge inference, and computer-vision components that turn a fleet of connected devices into a clinical fabric. Contact us if you have a specific IoMT integration or edge-AI question we can help unpick. References admin (2020) ‘Coca-Cola Vending Machine’, Vertical Innovations Ltd., 13 April (Accessed: 19 February 2024). How AI Helps Physicians Improve Telehealth Patient Care in Real-Time telemedicine.arizona.edu (no date) (Accessed: 25 January 2024). Pros & Cons of Continuous Glucose Monitors for Young Children with Type 1 Diabetes (2019) UMass Chan Medical School. ‘Smart pills could “dumb down” medical care - EPR’ (no date) European Pharmaceutical Review. ‘St George’s University Introduces VR Surgical Training’ (2019) vStream Digital Media, 28 May (Accessed: 12 March 2024). The Guardian (2018) ‘I was a doctor prone to fainting. This is how I got over it’, 16 January (Accessed: 12 March 2024).
