Digitization Accuracy Report
Designed to document chart-element detection, coordinate mapping, and measurement extraction against known ground-truth values.
Research and validation
A methodology for testing the promise—not assuming it.
The June 2026 SRS defines an academic prototype, planned studies, and performance targets. It does not report completed model or clinical validation findings.
Formal deliverables
Four outputs connect engineering to evidence.
Clinical Validation Study
Designed to evaluate clinical relevance, review workflows, and the safe interpretation of digitized trends and anomaly flags.
User Manual for Parents
Parent-facing guidance written at or below the specified fifth-grade comprehension level with accessible, active instructions.
API Integration Guide
Technical guidance for clinic partners and potential future exchange with EMR systems through the API.
Dataset creation
Controlled charts, realistic capture variation.
The proposed training dataset uses printed Kenya MOH Road-to-Health chart pages filled by volunteers. It does not use real patient charts in this academic prototype.
Why this matters: avoiding real patient charts removes the immediate need for hospital ethics approval during prototype dataset construction, while known plotted coordinates provide measurable ground truth.
- Print official chart pages
Use separate weight-for-age and height-for-age pages for boys and girls.
- Assign known coordinates
Give volunteers pre-defined age and measurement values to plot.
- Vary handwriting marks
Use blue or black ballpoint dots, crosses, or short lines.
- Vary devices and light
Photograph each chart by two people, on two devices, under at least two lighting conditions.
- Add real-world wear
Intentionally crease, fold, or lightly stain a subset.
- Review image quality
The team lead checks images before annotation.
Annotation and model training
A held-out test set protects the final measurement.
Roboflow is specified for annotation. The final test images must be photographed by people who did not fill the training charts.
Volunteer-filled charts
Known plot values and varied physical conditions.
→Chart classes
Axes, percentile curves, doctor plot points, and related noise classes.
→80% train
Cloud GPU training is planned because the team has no local GPU workstation.
→10% validation
Used during model development and hyperparameter selection.
10% test
Held out from training and photographed by separate participants.
→Ground truth
Compare detections and extracted values with the original assigned coordinates.
→Accuracy report
Document method, errors, limits, and performance.
→Validation study
Assess meaning and workflow beyond technical accuracy.
Specified performance targets
Targets for testing—not published results.
The SRS defines thresholds the implementation is expected to meet. Evidence must come from the held-out test set and clinical comparison work.
- 98% classification accuracy
Minimum target for bounding boxes and standard orientation markers across varied lighting.
- mAP 0.92
Target stated for the custom YOLOv11 chart-element pipeline.
- Less than ±50 g
Target extraction error tolerance for weight readings compared with manual assessment.
- Less than ±0.5 cm
Target extraction error tolerance for infant height readings.
- 99.9% monthly availability
Target for the future FastAPI and PostgreSQL cloud deployment.
WHO LMS calculations
Reference calculations need independent verification too.
The analytics engine is specified to select sex-specific WHO reference tables by exact age in days and calculate a Z-score using the LMS formula.
Z = ((Y / M)L − 1) / (L × S)
Y is the observed measurement; M is the median; L is the Box-Cox power; S is the coefficient of variation.
Validation questions
- Reference selection
Does the system choose the correct table for sex, measure, and exact age?
- Formula parity
Do outputs match a trusted independent WHO calculation?
- Extended range
Are values beyond ±3 SD handled with the specified extended formula?
- Display mapping
Are Z-scores translated into the right labelled percentile bands?
Research limitations
What the prototype cannot establish on its own.
One chart format
The model is constrained to the Kenya MOH Road-to-Health chart geometry. Results cannot be generalized to other chart formats without recalibration and testing.
Synthetic chart filling
Volunteer-filled charts can model handwriting and physical wear, but they do not capture every condition present in real clinical records.
No completed clinical findings
The SRS proposes a Clinical Validation Study but provides no outcomes. Clinical safety and workflow value remain to be demonstrated.
Academic prototype timeframe
The project is constrained to a semester prototype. Full clinical deployment is explicitly outside the stated scope.
Cloud training dependency
Model training depends on external cloud GPU resources because the team does not have a local GPU workstation.
Scope inconsistency
The project scope calls immunization tracking out of scope, while Module 8 defines it as an MVP requirement. This needs formal requirements reconciliation before validation.
Research integrity
NeoScan should earn trust through transparent evidence.
Future reports should publish the dataset protocol, split logic, evaluation definitions, error distribution, device and lighting coverage, failure cases, manual override rate, clinical review method, and limitations—not only one headline metric.
Research collaboration