We built a streaming ML system for a hospital ICU that reads monitor and lab data and sends explained sepsis alerts to nurses. It flagged sepsis 4.1 hours earlier and was linked to an 18% drop in sepsis deaths.
Client: NextGen PharmaIndustry: HealthcareService: AI & Automation
Overview
Client
NextGen Pharma
Hospital intensive care unit
Engagement snapshot
The challenge
The hospital's ICU relied on separate monitor alerts and fixed EMR thresholds, which often fired only after the body's inflammatory response was well under way. Every hour of delay in sepsis treatment increases mortality risk by 8%, yet staff were swamped by false monitor alarms.
Changes in vital signs were often spotted only on nursing rounds. Most device alerts were irrelevant, and blood pressure, heart rate and lactate sat in separate systems, so no one saw the full picture of a patient's decline.
Important changes in vital signs were often found only during 4-hour nursing rounds, missing the "golden hour" for treatment.
Over 70% of device alerts were clinically irrelevant, so ICU nurses had learned to tune them out.
Blood pressure, heart rate and lactate levels were processed in different systems, which hid a patient's overall decline.
The solution
We combined streaming data from bedside monitors (HL7) and lab results from the health record into one fixed schema, so the model's inputs stay consistent in live use.
A gradient-boosted model runs on a cloud streaming engine and looks for unusual patterns across 12 vital signs at once.
An automated clinical agent sends prioritised alerts straight to nurse workstations, each with SHAP reason codes that explain why the patient was flagged.
Key features
A model scores 12 vital signs together in real time to spot early sepsis patterns.
Business impact
Sepsis flagged 4.1 hours earlier
Nurses get prioritised alerts with plain reason codes, and noisy spikes are filtered out.
Business impact
65% fewer false alarms
Calculates rolling averages and rates of change for each patient's vital signs, such as mean arterial pressure (MAP) and oxygen saturation (SpO2), to filter out noisy monitor spikes.
Business impact
65% fewer false alarms
Models score de-identified vital-sign streams in memory. Patient identity is re-linked only at the secure clinical endpoint.
Business impact
Built to HIPAA requirements
Alerts are escalated automatically based on unit staffing levels and how ill each patient is.
Business impact
18% lower mortality
Before and after
Before
After onset
Alerts fired only once white cell counts or blood pressure crossed critical thresholds.
After
4-hour lead
Pattern recognition spots a patient's decline before clinical thresholds are crossed.
Before
Manual chart entries
Analysis depended on nurses' chart entries, which are often late during busy shifts.
After
Direct from the monitors
Readings come straight from the monitors, so 100% of vital signs are captured with no wait for manual charting.
Before
One bed at a time
Each bedside device worked on its own, with no analysis across units.
After
Cloud platform
The system is built to monitor 500+ ICU beds at once and scores each reading in under 100ms.
Results
4.1h Detection Lead
The streaming model spotted signs of sepsis 4.1 hours earlier than the hospital's previous manual process.
Before: ReactiveAfter: 4h Early
65% Lower False Alarm Rate
Scoring several vital signs together filtered out random monitor noise, so nurses could focus on the patients who needed them.
Before: NoisyAfter: Filtered
18% Lower Mortality Rate
Earlier treatment was linked to an 18% reduction in sepsis-related deaths in the trial ICU units.
Before: StandardAfter: 18% Lower
Testimonial
“Our nurses now get sepsis alerts 4.1 hours earlier, with plain reason codes attached. False alarms are down 65% and mortality is 18% lower. For an intensive care unit, those are the numbers that count.”
Dr. Claire HennesseyICU Medical Director, NextGen PharmaKey results
Trust and control
Each alert shows SHAP reason codes, such as rising heart rate plus falling SpO2, so clinicians can see why it fired.
Compliance readyAll workflows run inside the hospital's private cloud, and no patient health information leaves that secure environment.
Built to HIPAA requirementsBackup cloud services keep the warning system running even during EHR or hospital Wi-Fi outages.
Cloud nativeThe hospital owns 100% of the sepsis models and alerting logic once the project is complete.
100% ownedFaster delivery
These modules saved about 7 weeks of setup, which went into tuning the sepsis model and alert rules with the clinical team.
Gave us an encrypted, access-controlled base for the patient data streams, built to HIPAA requirements.
Provided tested handling for bedside monitor feeds, including the rolling averages used to filter out noise.
Shows nurses and clinical leads the risk scores and alerts for every monitored bed.
Kept warnings running near the bedside when the health record system or hospital Wi-Fi went down.
The system streams data from bedside monitors and lab results into one schema, and a gradient-boosted model looks for unusual patterns across 12 vital signs at once. Each reading is scored in under 100ms. This lets it spot a patient's decline before standard thresholds are crossed, 4.1 hours earlier than before.
The new alerts cut false alarms by 65%. Before, over 70% of device alerts were clinically irrelevant. The system now calculates rolling averages and rates of change for each patient, which filters out noisy monitor spikes, and every alert shows SHAP reason codes that explain why the patient was flagged.
All workflows run inside the hospital's private cloud, built to HIPAA requirements, and no patient health information leaves that environment. Models score de-identified vital-sign streams in memory, and identity is re-linked only at the secure clinical endpoint. Data is encrypted with AES-256, and backup services keep alerts running during outages.
If your clinical teams are buried in monitor alarms but still spot deterioration late, we can help you build early warnings they can trust.