Validated on 4.5 million adults

Redefining "Normal"
in Laboratory Medicine

NORMA personalizes lab interpretation using a patient's history, demographics, and clinical state — detecting meaningful change earlier than static reference ranges.

See a live example Try NORMA
4.5M
adults validated
2.8M
training sequences
31
biomarkers
1.7B
lab results
The Problem

Normal for whom?

A single reference range cannot capture individual biology. Clinically meaningful change often hides in plain sight — inside "normal."

Clinical Example

A patient's fasting glucose rises gradually but remains within population "normal" for months. NORMA flags the deviation 6 months before the standard reference range.

Problem
Static reference ranges miss early disease by treating every patient's biology as interchangeable.
Solution
NORMA learns what’s normal for each patient from longitudinal history and clinical context.
Impact
Earlier detection across hundreds of conditions — validated on 4.5M patients at national scale.
Conditions detected earlier with personalized ranges
Anemia Diabetes Kidney disease Liver disease Dyslipidemia Thyroid disorders Infection Bleeding disorders
Why It Matters

Reference ranges shape every clinical decision

A single lab result classified as "normal" or "abnormal" triggers a cascade of downstream actions — from treatment plans to insurance coverage.

Lab Result Normal or Abnormal? Reassure Monitor Treat / Refer
Clinical Decisions
Treatment plans and specialist referrals
Whether a doctor reassures, orders follow-up testing, adjusts medication, or refers to a specialist depends on where a value falls relative to the reference range.
Normal Normal Drifting Flagged Diagnosis time → missed window
Disease Risk
Early detection vs. missed progression
A value drifting toward disease may stay "normal" by population standards for months or years. The wrong reference range delays detection and narrows the intervention window.
LAB REPORT Approved Coverage / lower premiums Denied Higher premiums / exclusions "Normal" "Abnormal"
Financial Decisions
Insurance eligibility and coverage
Lab classifications directly affect insurance underwriting, premium rates, and whether procedures or treatments are covered. A misclassified "abnormal" can have lasting financial consequences.
Framework

Three approaches to laboratory interpretation

Population Reference
Too Generalized
  • Fixed thresholds from a reference population
  • Ignores individual physiological baseline
  • Meaningful change may go undetected within range
  • Sensitivity limited by inter-individual variation
Standard of care in clinical laboratory medicine
Pure Individualization
Too Personalized
  • Fitted to the patient's own history only
  • Overfits on sparse or noisy observations
  • May normalize chronic disease states
  • No population anchor to constrain predictions
Prior work on individual setpoint models
NORMA
Contextualized
  • Trained on population-scale longitudinal sequences
  • Conditioned on individual trajectory and demographics
  • Predicts under a specified clinical state
  • Calibrated uncertainty adapts to data and horizon
  • Externally validated without retraining
This work
Coverage — 31 biomarkers

Hover a biomarker to see its clinical context and reference intervals.

A1C
ALB
ALP
ALT
AST
BUN
CA
CL
CO2
CRE
DBIL
GLU
HCT
HDL
HGB
K
LDL
MCH
MCHC
MCV
MPV
NA
PLT
PT
RBC
RDW
TBIL
TC
TGL
TP
WBC
Training Cohorts

Trained on two health systems. Validated in a third.

Longitudinal lab sequences from MIMIC-IV and EHRSHOT. External validation on 4.5M adults in Clalit Health Services — no retraining.

4.5M
adults validated
34
biomarkers
1.7B
lab results
Training cohort
External validation
CohortInstitutionRoleSequencesSplit (70 / 10 / 20)
MIMIC-IV Beth Israel Deaconess Medical Center, Boston MA Training 1,907,032
EHRSHOT Stanford Medicine, Palo Alto CA Training 96,042
Clalit Clalit Health Services, Israel External validation 4,500,000+ Held-out — no retraining
Forecasting Performance

How well does NORMA predict?

Forecasting accuracy — how closely NORMA's predicted distribution matches observed next values. High R² means the model captures individual trajectory dynamics well.

Train
0.762
R²  ·  MAE 5.97
Validation
0.75
R²  ·  MAE 5.7
Test (held-out)
0.747
R²  ·  MAE 5.73
Bootstrap estimates — split:
Loading metrics…
Clinical Validation

Personalized ranges predict disease — but only when anchored to population trends

Clinically validated on 4.5 million individuals in Clalit Health Services, Israel. These charts show cases that population reference ranges missed entirely — over-personalization catches them, but floods clinicians with false positives.

NORMA detects early disease without the noise.
More cases caught Fewer false alarms More reliable flags Personalized was better
Mortality
Chronic Kidney Disease
Type 2 Diabetes

Each wedge = one lab test. Bar height = NORMA's improvement over purely personalized ranges. Center = largest per-test gain. Hover for details.

How it works

NORMA — Normal Outcome Range Modeling with Attention

A causal transformer that learns patient-specific lab dynamics and outputs calibrated predictions conditioned on clinical state.

1
Patient History
Input features
Lab values
Numeric measurement at each visit
Clinical state
Low / Normal / High per time step
Time intervals
Irregular gaps encoded via Time2Vec
Age & sex
Demographic context
Lab test ID
Which biomarker
Query token — specifies target state & forecast horizon
2
Causal Transformer
Encoder
8
Layers
4
Heads
64
Dim
565K
Params
Causal masking ensures each step attends only to past observations — no future leakage.
Last hidden state encodes the full trajectory, conditioned on target state and horizon.
3
Prediction
Output head
μ
Predicted mean
Expected next lab value
σ
Predicted uncertainty
Narrows with data; widens over time
Together: a 95% predictive interval personalized to each patient
Gaussian NLL Loss
ℒ = ½ [ log σ² + (y − μ)² / σ² ]
Accuracy (y − μ)² · Calibration log σ²
Healthy-state query
"What value would we expect if this patient were healthy?"
Set state = Normal at inference. Deviation from this expectation is a clinically meaningful signal.
What-if query
"What if this patient were older, or a different sex?"
Swap demographic inputs at inference. The model adapts its prediction to the new context.

See it in action.

Enter a lab history and see how NORMA classifies each value differently than population ranges.

See a live example Try your own data