Wearables make it possible to continuously monitor patients outside the bedside, but the clinical value does not lie in collecting more data: it lies in transforming reliable time series into actionable alerts. A safe system combines appropriate sensors, clinical context, trend analysis, electronic health record integration, and protocols that define who will be notified, within what timeframe, and what course of action should be followed.
What continuous telemetry with wearables is
Continuous telemetry is the recurring collection of physiological and behavioral signals through wearable devices or connected sensors. Depending on the equipment, heart rate, peripheral oxygen saturation, respiratory rate, temperature, blood pressure, electrocardiogram, sleep, physical activity, and falls can be monitored.
“Continuous” does not necessarily mean transmitting every sample in real time. The architecture can operate in three ways:
- Real time: data is sent within seconds, which is appropriate when delays may affect the clinical response.
- Near real time: synchronization occurs at intervals of a few minutes, balancing urgency, battery life, and connectivity.
- Batch processing: periodic transmission, useful for longitudinal monitoring without the need for immediate intervention.
The appropriate frequency depends on the patient’s risk and clinical usefulness. Transmitting heartbeats every second may be justified in a monitored unit, but unnecessary for tracking physical activity after a stable discharge.
How predictive alerts work
Traditional alerts compare a measurement against a fixed threshold: for example, they signal when a specific indicator falls outside a range configured by the care team. Predictive alerts analyze the combined evolution of data to estimate deterioration before an isolated threshold is exceeded.
A typical pipeline has five stages:
- Acquisition: a wearable, gateway, or application captures and transmits the signals.
- Validation: the system identifies missing data, impossible values, artifacts, and connection problems.
- Contextualization: information is associated with the patient, diagnosis, medication, procedure, and individual baseline.
- Inference: clinical rules or machine learning models calculate risk and trends.
- Orchestration: an alert is sent to the appropriate professional, including its priority, rationale, and response deadline.
AI can identify relationships among variables that would be difficult to assess manually. However, its output must be presented as decision support, not as an autonomous diagnosis. A score without an explanation, context, or response protocol tends to increase the team’s workload rather than reduce it.
Fixed thresholds, trends, or AI models?
Each approach addresses a different level of complexity:
- Fixed thresholds: simple, auditable, and fast, but not highly personalized and susceptible to false positives.
- Trends and composite rules: consider temporal variation and multiple signals; they require more careful clinical design.
- Predictive models: can recognize complex patterns, but require representative data, external validation, and performance monitoring.
In practice, a hybrid architecture is usually safer. Deterministic rules protect against known critical events, while predictive models prioritize cases and identify subtle changes. The decision should not be whether or not to use AI, but which combination provides useful sensitivity without making the operation unfeasible.
Data quality comes before the algorithm
Wearables are subject to movement, sweat, incorrect positioning, differences in skin pigmentation, peripheral perfusion, low battery, and loss of connection. General-purpose equipment may also have a different intended use from validated medical devices.
Before feeding data into a model, the platform must record at least:
- the device’s source, manufacturer, and model;
- the measurement’s unit, frequency, and time;
- the quality or confidence reported by the sensor;
- periods of non-use and loss of connectivity;
- transformations performed on the original data;
- clock and time zone synchronization;
- an unambiguous link among the device, patient, and care episode.
An abnormal reading with poor signal quality should not be treated in the same way as a persistent trend confirmed by multiple measurements. Instead of hiding uncertainty, the system should display it and, when necessary, request a new measurement or confirmation using clinical equipment.
How to prevent alert fatigue
Alert fatigue occurs when professionals receive so many irrelevant notifications that they begin to ignore them or take longer to respond. The problem is not limited to the interface: it involves clinical criteria, data quality, team staffing, and governance.
Practical measures include:
- combining absolute value, duration, and rate of change;
- requiring persistence over a defined window before issuing an alert;
- suppressing repetitions during a controlled interval;
- grouping related signals into a single event;
- adapting thresholds to the individual baseline when clinically valid;
- classifying priority according to risk and expected response time;
- automatically escalating unacknowledged alerts;
- recording acknowledgment, course of action, and operational outcome.
Four metrics help evaluate the system: alerts per patient/day, proportion acknowledged, time to acknowledgment, and proportion that resulted in clinical action. Model sensitivity and specificity also matter, but they do not show by themselves whether the workflow functions in the real world.
Integrating with electronic health records using HL7 v2 and FHIR
An isolated platform creates one more screen for the team to check. Integration should bring observations and alerts into the electronic health record or care system used in the daily workflow.
In legacy environments, HL7 v2 remains relevant for exchanging admission, discharge, transfer, and results messages. FHIR provides resources and APIs that are better suited to modern applications. Depending on the use case, measurements can be represented by Observation, devices by Device, patients by Patient, and episodes by Encounter.
Semantic interoperability requires more than transporting JSON or HL7 messages. Units, codes, identifiers, times, provenance, and clinical meaning must be standardized. Mechanisms for idempotency, patient reconciliation, reprocessing, and auditing must also be in place.
Security, LGPD, and clinical responsibility
Health data is sensitive personal data under the Brazilian General Data Protection Law. The project must define its purpose, applicable legal basis, retention, sharing, data subject rights, and responsibilities among the hospital, platform, manufacturers, and operators.
Minimum technical controls include encryption in transit and at rest, strong authentication, role-based access, segregation between environments, audit trails, vulnerability management, and an incident response plan. Applications, APIs, gateways, and devices must be included in the threat model; protecting only the database is insufficient.
When software and devices influence care decisions, it is also necessary to verify the applicable regulatory classification and Anvisa requirements. This analysis depends on the intended use, functionalities, and product claims, and should involve clinical, legal, and regulatory stakeholders.
Criteria for deciding whether the project makes sense
Remote monitoring tends to generate value when there is a well-defined population, a relevant clinical event, and a team capable of taking action. Before development begins, answer:
- What deterioration or event should be identified?
- How much advance notice would be clinically useful?
- Which signals are available with sufficient quality?
- Who will receive the alert, and through which channel?
- What will the acknowledgment and escalation timeframe be?
- What action can be taken after the alert?
- How will false positives and false negatives be measured?
- How will performance be compared across patient profiles?
- Does the wearable have compatible battery life, connectivity, and usability?
- Is there integration with identification, admission, and electronic health record systems?
If no action can be taken, the alert has little value, regardless of how sophisticated the algorithm may be.
Implementation in stages
A responsible implementation can be divided into four phases.
1. Clinical and technical definition
Select a specific care journey, such as post-discharge follow-up or chronic patient monitoring. Map signals, devices, risks, responsible parties, and indicators before choosing the AI model.
2. Silent pilot
Run the algorithm without sending alerts to the team. Compare predictions with observed events and evaluate missing data, latency, subgroups, and the potential volume of notifications.
3. Assisted operation
Enable alerts for a limited group, with supervision and an explicit protocol. Collect feedback from professionals and record the causes of alerts that provide no value.
4. Scaling and monitoring
Expand only after meeting previously approved criteria. Monitor model degradation, device changes, protocol updates, incidents, and the impact on care delivery.
How Predictor Solutions addresses this
Predictor Solutions develops healthcare systems with HL7 v2 and FHIR integration, data pipelines, applied artificial intelligence, and cloud infrastructure with security controls. Predictor Health brings together health dashboards and wearable data; Predictor AI Hospitals works with the prediction of sepsis, myocardial infarction, and pneumonia in ICUs, always within workflows that require clinical validation and operational integration.
The approach combines clinical discovery, interoperable architecture, signal quality assessment, auditable alerts, and gradual implementation. The company, based in Lavras, Minas Gerais, also applies data engineering, DevOps, and offensive security to reduce risks when bringing models into production. Across its overall portfolio, Predictor Solutions serves 9 medium-sized and large companies, with reported average results of R$ 1.32 million in savings per client per year, a 70% increase in productivity, and a 43% increase in profit within six months; these figures do not represent a promise of clinical outcomes and depend on the context of each project.
Contact: contato@predictorsolutions.com / WhatsApp +55 31 98835-3246.