AnemiaAssist
CBC clinical decision support
● Loading model🔒 Browser-onlyNo patient data transmitted
Clinician-facing research prototype

Structured anemia assessment from CBC data—without sending data to a server.

Combines WHO 2024 haemoglobin categories, a calibrated morphology-risk ensemble, uncertainty referral, conformal prediction, morphology heuristics and printable clinical review prompts.

1. Patient context

Used for WHO haemoglobin interpretation and scope checks. It is not an ML predictor.

Not stored after page refresh.

2. Core CBC measurements

All eight fields are required by the browser model. Verify units against the analyzer report.

10⁹/L
10¹²/L
g/dL
%
fL
pg
g/dL
10⁹/L
Important clinical boundaryThe model predicts similarity to morphology-derived target labels in the development cohort. Anaemia diagnosis and severity must be based on validated haemoglobin measurement, patient context, laboratory reference intervals and clinician judgment.

No assessment yet

Enter a patient context and eight CBC measurements, then select Analyze CBC.

Batch CSV analyzer

Client-side processing for de-identified datasets. No file is uploaded.

Required columns: WBC,RBC,HGB,HCT,MCV,MCH,MCHC,PLT. Optional: patient_id,population,adjusted.

Model architecture and evidence

Versioned browser model exported from the corrected 744-record CBC–morphology dataset.

—

Architecture

Core CBC → engineered indices → logistic component + gradient-boosting component → validation-derived weighted blend → Platt calibration → clinical threshold + uncertainty referral + class-conditional conformal set

Model target

Probability that the CBC pattern is consistent with an abnormal/anemic RBC morphology report in the research dataset. It is not the same as a clinical diagnosis of anaemia or its cause.

Internal validation

Five-fold cross-fitted base predictions, cross-fitted calibration, 1,000 stratified bootstrap resamples, and thresholds selected without a separate patient input.

Known limitations

  • No age, sex, pregnancy, ethnicity, analyzer, site or altitude metadata were available for model development.
  • Labels were generated from paired morphology report wording and require expert adjudication.
  • No external or prospective clinical validation has been performed.
  • Probability estimates may require recalibration in a new laboratory or patient population.
  • The model may underperform in normocytic abnormal patterns and unusual aetiologies.

Safety, privacy and governance

Privacy by design

  • All calculations run in the browser.
  • No API, analytics service, database or external CDN is used.
  • Patient values are not placed in the URL.
  • Only the disclaimer acknowledgement is stored locally.

Required local governance

  • Validate against your laboratory population and analyzer.
  • Define clinical owners, change control and incident reporting.
  • Review WHO cutoff adjustments and local critical-value policies.
  • Audit subgroup performance and referral outcomes.
Do not use as an autonomous diagnostic systemThis prototype must remain under qualified clinician supervision. Do not use it to prescribe treatment, delay urgent care, or override laboratory/clinical findings.

References included in the implementation

  • World Health Organization. Guideline on haemoglobin cutoffs to define anaemia in individuals and populations. 2024.
  • GitHub Pages official documentation for static site deployment.