Patient Deterioration Prediction
“Which inpatients will deteriorate in the next 12 hours?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which inpatients will deteriorate in the next 12 hours?
Failure to rescue costs U.S. hospitals $1.1B annually. Traditional early warning scores (NEWS, MEWS) generate false alarm rates above 90%, causing alert fatigue. A 600-bed hospital averages 15 unexpected ICU transfers per month; each costs $31,000 more than a planned transfer. Catching deterioration 6 hours earlier would save $5.6M per year and 40+ lives.
Quick answer
AI predicts inpatient deterioration by fusing vital sign trajectories, lab trends, medication administrations, and nursing assessment patterns into a real-time relational graph. Traditional early warning scores (NEWS, MEWS) use single-point-in-time vital thresholds and generate false alarm rates above 90%, causing alert fatigue. Graph-based models detect deterioration 6 hours earlier with 80% fewer false alarms by learning temporal patterns across multiple data streams and cross-patient signals from the same unit.
Approaches compared
4 ways to solve this problem
1. NEWS/MEWS (National Early Warning Score)
Aggregate scores based on vital sign thresholds: heart rate, blood pressure, respiratory rate, temperature, SpO2, and consciousness level. Calculated at each nursing assessment. Escalation triggers at score thresholds.
Best for
Universal screening where every patient needs a baseline risk assessment. Simple to calculate bedside. Supported by extensive clinical validation.
Watch out for
Single-point-in-time scoring misses trajectories. A heart rate of 102 scores the same whether it was 72 four hours ago (concerning trend) or 108 four hours ago (improving trend). False alarm rates above 90% cause alert fatigue, and nurses learn to ignore the warnings.
2. EHR-Embedded Sepsis Alerts
Vendor-provided sepsis screening tools (Epic Sepsis Model, Cerner St. John Sepsis) that monitor for SIRS criteria, qSOFA, and sepsis-related lab patterns. Integrated into clinical workflow with automated alerts.
Best for
Sepsis-specific detection where early antibiotics save lives. Embedded in the nursing workflow without additional tools.
Watch out for
Sepsis-specific: misses cardiac, respiratory, neurological, and post-surgical deterioration entirely. High false-positive rates (90%+ in published studies of the Epic Sepsis Model) reduce clinical trust.
3. Single-Patient Time-Series Models
Deep learning models (LSTMs, transformers) trained on individual patient vital-sign and lab-result time series to predict deterioration within 4-12 hours. Captures temporal trends within a single patient's data.
Best for
Patients with rich temporal data streams (ICU step-downs, telemetry patients) where the trajectory within the patient's own data is the primary signal.
Watch out for
Operates on one patient at a time. Cannot detect unit-level signals: when multiple patients on the same floor show subtle vital changes simultaneously, it may indicate a staffing problem, equipment issue, or medication error that affects the whole unit.
4. Graph Neural Networks (Kumo's Approach)
Fuses vitals trajectories, lab trends, medication administrations, nursing assessments, and unit-level context into a temporal relational graph. Detects both individual patient deterioration patterns and cross-patient signals on the same unit.
Best for
Catching deterioration 6-8 hours before traditional scores trigger, with 80% fewer false alarms. Detects unit-level risk factors (understaffing, concurrent deterioration) that affect individual patient outcomes.
Watch out for
Requires real-time data feeds from vitals monitors, labs, and medication administration records. Latency in data availability directly impacts prediction lead time.
Key metric: NEWS/MEWS generate 90%+ false alarms. Graph-based models achieve 80% fewer false alarms while detecting deterioration 6 hours earlier, saving $5.6M annually and 40+ lives at a 600-bed hospital.
Why relational data changes the answer
Flat deterioration models analyze each patient's vitals independently: is heart rate above threshold? Is SpO2 below threshold? They catch the patient whose vitals have already crossed into dangerous territory. But they miss the patient whose MAP has been declining slowly for 4 hours, who just received a new vasopressor (indicating the care team is already concerned), whose lactate came back elevated, and whose nursing note mentions 'increased confusion.' These signals span four different data streams (vitals, medications, labs, notes), and their temporal co-occurrence is what predicts deterioration 6-8 hours before any single vital sign crosses a threshold. A flat model would need each of these as a pre-engineered feature, and it would still miss the timing relationships between them.
Relational learning connects these data streams temporally. The model tracks the trajectory of vitals over sliding windows, correlates them with medication changes (a new vasopressor following a blood pressure drop is a strong signal), lab trends (rising lactate concurrent with vital changes confirms the pattern), and nursing sentiment (concern documented before vitals cross thresholds). It also detects cross-patient signals: when 3 patients on the same Med-Surg unit show subtle vital changes within the same 2-hour window, the model weighs unit-level risk factors (staffing ratio, recent shift change) that may explain the concurrent deterioration. This multi-stream, multi-patient approach produces a C-statistic above 0.90 with a false alarm rate below 15%, compared to NEWS/MEWS at 0.70-0.75 with false alarms above 90%.
Detecting patient deterioration from single-point vitals is like monitoring a building for structural failure by checking one sensor at a time. The temperature sensor says 'normal,' the vibration sensor says 'normal,' the moisture sensor says 'normal.' But if you connect them, you see that temperature is rising slowly, vibration shifted 2 hours ago, and moisture appeared at the foundation yesterday. The building is not failing yet, but the pattern across sensors says it will. Relational deterioration models are the connected building-monitoring system for hospitalized patients.
How KumoRFM solves this
Graph-learned clinical intelligence across your entire patient network
Kumo builds a temporal graph from vitals streams, lab results, medication administrations, and nursing assessments. It learns that a specific pattern of trending MAP decline combined with a new vasopressor order and a nursing note sentiment shift precedes deterioration 8 hours before traditional scores trigger. The model weighs cross-patient signals: when multiple patients on the same unit show subtle vital changes simultaneously, it detects environmental or staffing-related risk factors.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
PATIENTS
| patient_id | age | unit | admit_date |
|---|---|---|---|
| P6001 | 71 | Med-Surg 4A | 2025-03-01 |
| P6002 | 55 | Med-Surg 4A | 2025-03-02 |
| P6003 | 83 | Telemetry 3B | 2025-02-28 |
VITALS
| vitals_id | patient_id | timestamp | hr | bp_sys | spo2 | temp |
|---|---|---|---|---|---|---|
| VT01 | P6001 | 2025-03-03 06:00 | 88 | 132 | 96 | 37.1 |
| VT02 | P6001 | 2025-03-03 10:00 | 102 | 108 | 93 | 38.2 |
| VT03 | P6002 | 2025-03-03 08:00 | 72 | 128 | 98 | 36.8 |
LABS
| lab_id | patient_id | test_name | value | timestamp |
|---|---|---|---|---|
| LB01 | P6001 | Lactate | 3.8 | 2025-03-03 07:00 |
| LB02 | P6001 | WBC | 16.4 | 2025-03-03 07:00 |
| LB03 | P6002 | Lactate | 1.1 | 2025-03-03 06:00 |
MEDICATIONS
| med_id | patient_id | drug | route | admin_time |
|---|---|---|---|---|
| MD01 | P6001 | Norepinephrine | IV | 2025-03-03 09:30 |
| MD02 | P6002 | Metoprolol | PO | 2025-03-03 08:00 |
| MD03 | P6003 | Heparin drip | IV | 2025-03-03 06:00 |
NURSING_NOTES
| note_id | patient_id | timestamp | assessment |
|---|---|---|---|
| NN01 | P6001 | 2025-03-03 08:00 | Pt appears more confused, family concerned |
| NN02 | P6002 | 2025-03-03 08:30 | Resting comfortably, no complaints |
| NN03 | P6003 | 2025-03-03 07:00 | Steady, ambulating in hallway |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(VITALS.RAPID_RESPONSE, 0, 12, hours) FOR EACH PATIENTS.PATIENT_ID WHERE PATIENTS.UNIT != 'ICU'
Prediction output
Every entity gets a score, updated continuously
| PATIENT_ID | UNIT | TIMESTAMP | DETERIORATION_PROB |
|---|---|---|---|
| P6001 | Med-Surg 4A | 2025-03-03 10:15 | 0.91 |
| P6002 | Med-Surg 4A | 2025-03-03 10:15 | 0.07 |
| P6003 | Telemetry 3B | 2025-03-03 10:15 | 0.15 |
Understand why
Every prediction includes feature attributions — no black boxes
Patient P6001 -- 71yo, Med-Surg 4A
Predicted: 91% deterioration probability in next 12 hours
Top contributing features
Heart rate trend (last 4h)
+14 bpm
27% attribution
Lactate level
3.8 mmol/L
24% attribution
New vasopressor initiated
Norepinephrine
21% attribution
SpO2 decline rate
-3% in 4h
16% attribution
Nursing assessment sentiment
Concern noted
12% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about patient deterioration prediction
How does AI predict patient deterioration in hospitals?
AI predicts deterioration by fusing multiple real-time data streams: vital sign trajectories, lab result trends, medication administration patterns, and nursing assessment narratives. Graph-based models learn that specific temporal patterns across these streams (e.g., declining MAP + new vasopressor + rising lactate + nursing concern) predict deterioration 6-8 hours before traditional early warning scores trigger.
What is the false alarm rate for early warning scores like NEWS?
Traditional early warning scores (NEWS, MEWS) generate false alarm rates above 90% in most published studies. This means more than 9 out of 10 alerts are false positives, causing severe alert fatigue. Nurses learn to dismiss warnings, and real deterioration events get lost in the noise. Graph-based models reduce false alarms to below 15% while detecting deterioration earlier.
How many hours in advance can AI detect patient deterioration?
Graph-based models detect deterioration 6-8 hours before traditional threshold-based scores trigger alerts. The lead time comes from tracking vital-sign trajectories and correlating them with medication changes and lab trends. A declining MAP that has not yet crossed a threshold is invisible to NEWS/MEWS but is captured by trajectory-aware models.
What is the cost of failure to rescue in hospitals?
Failure to rescue (death following a complication that could have been treated if detected earlier) costs US hospitals $1.1B annually. Each unexpected ICU transfer costs $31,000 more than a planned transfer. A 600-bed hospital averages 15 unexpected ICU transfers per month. Catching deterioration 6 hours earlier prevents 40+ deaths and saves $5.6M annually.
Bottom line: A 600-bed hospital catching inpatient deterioration 6 hours earlier prevents 40+ deaths and saves $5.6M annually in unplanned ICU transfers. Kumo fuses vitals trajectories, lab trends, medication signals, and nursing narrative patterns into a single real-time score that outperforms NEWS/MEWS with 80% fewer false alarms.
Related use cases
Explore more healthcare use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
