We built an AI platform for a regional health system that filters noise from wearable data and alerts clinicians early to heart failure risk. Readmissions fell 18%, and clinicians intervened 40% faster.
Industry: HealthcareService: AI & AutomationUpdated
Overview
The challenge
The client's remote monitoring programme produced millions of data points a day from patient wearables. Clinicians were swamped by false high-heart-rate alerts set off by everyday things like walking or stress.
Nurses spent hours sorting non-urgent alerts while real warning signs were missed, which led to costly emergency admissions. The older EMR could not handle streaming data, so physicians lacked live context.
Nurses had to sort 5,000+ non-urgent alerts a week by hand, which slowed their response to high-risk patients.
Missing early fluid retention or arrhythmia led to emergency hospital stays costing $30k+ each.
The older EMR systems could not process high-frequency streaming data, so physicians had no live context.
The solution
We automated a secure connection that takes encrypted data from connected medical devices and removes identifying details at the edge, before any clinical scoring.
AI clinical tools compare heart rate, sudden weight gain and oxygen levels to judge whether a patient's heart failure is really getting worse.
Every alert is saved with the physiological trends and clinical reasoning behind it, so physicians can trace each step to the intervention.
Before and after
Before
Fixed thresholds
Simple high and low limits caused constant false alarms whenever patients moved.
After
AI with context
Data from several sensors is compared to tell exercise apart from an emergency.
Before
Perimeter trust
Data sat behind VPNs and was exposed if the main node was breached.
After
Zero-knowledge platform
End-to-end encryption, with patient identity removed before clinical scoring.
Before
Manual dashboard checks
Clinicians checked dashboards by hand, and responses lagged by hours or days.
After
Automatic trigger
The AI starts a preventive intervention workflow in under 60 seconds.
Key features
The AI spots early trends, such as gradual weight gain alongside falling heart rate variability (HRV), days before heart failure symptoms appear.
Business impact
48-hour early detection
Models run on anonymised data inside a private cloud built to HIPAA requirements, so patient identities are never exposed outside it.
Business impact
Zero patient data leakage
Critical alerts go straight to the on-call cardiologist's device, with the full EMR context attached.
Business impact
40% faster intervention
Results
18% Admission Reduction
Early alerts let clinicians adjust diuretics in time, keeping patients out of the emergency room.
Before: High-RiskAfter: 18% Lower
40% Response Lift
Filtering out 92% of the noise let staff act on critical signals 40% faster.
Before: LaggedAfter: Instant
500ms p95 Response Time
The system handles 50k+ streams at once and processes readings in under a second.
Before: < 1sAfter: 500ms
Faster delivery
These modules saved about 8 weeks of setup, which went into the risk models and the EMR connection for live context.
Verified each connected medical device before its data was accepted.
Calculated the weight, heart rate and oxygen trends that the risk models score.
Shows care teams patient risk and open alerts across the 50,000+ monitored patients.
Checked encryption, access and logging against HIPAA requirements as features were added.
Trust and control
Patient data is encrypted under the client's business associate agreement (BAA), and personal details are kept separate across the system.
HIPAA safeguardsSecure, traceable logs record every path from sensor reading to alert, ready for malpractice insurance audits.
Verifiable dataThe cloud set-up adds capacity quickly during busy health events.
Ready to growThe health system owns 100% of the clinical agents and custom health record connections.
100% ownedInstead of simple high and low heart rate limits, the AI compares heart rate, weight changes and oxygen levels from several sensors. This tells everyday activity like walking apart from a real emergency. The platform filters out 92% of non-clinical noise, so nurses no longer sort 5,000+ non-urgent alerts a week.
Readmissions fell by 18%, because early alerts let clinicians adjust diuretics before patients needed emergency care. Clinicians intervened 40% faster, and the AI spots warning trends up to 48 hours early. Each avoided admission matters, since these emergency hospital stays cost $30k+ per event.
Identifying details are removed at the edge before any clinical scoring, and models run on anonymised data inside a private cloud built to HIPAA requirements. Data is encrypted end to end under the client's business associate agreement. Secure logs record every path from sensor reading to alert, ready for audits.
If wearable alerts are burying your clinicians in noise, we can help you filter it and send only the warnings that need action.