NIÑOSALUD AI   /   Release 1.2
← Back to dashboard   ·   Research note PDF
08 / RESEARCH EVIDENCE

What survived the checks?

Two separate exploratory analyses. Reproducible calculations, not prospective or clinical validation.

992

Matched EDA prediction rows
Eight provinces · 2023–2025 evaluation

Complexity did not beat persistence

MethodMAE / episodes
Persistence / last observed week45.088
Trailing four-week mean46.673
Random Forest + climate50.294
Random Forest without climate52.763
Previous-year same week71.954

Lower is better. A four-week-ahead target is one future week's count, not a four-week cumulative total. Training: 2019–2022.

+5.206

Climate forest minus persistence MAE

Uncertainty limits the conclusion

Approximate 95% cluster-bootstrap interval:
−0.474 to +13.349 episodes

5,000 resamples of 24 province-year sequences. The interval crosses zero; it is not an equivalence test or a prediction interval for future patients.

The evaluation years were inspected during development. Provinces and years may be correlated. No operational forecasting advantage has been demonstrated.

REGIONAL DENGUE / 2025 COHORT

The interpretation changes with the filter

Maximum retained intervalPiura median / daysHuancabamba median / days
30 days35
60 days310
90 days315
180 days318
365 days319

These are differences between recorded symptom-onset and attendance dates. Thresholds are analytical choices, not clinically certified validity rules. The source dictionary does not establish an independently verified measure of first access to care.

The 0–90-day view retains 5,468 records in Piura and 249 in Huancabamba. Repeating the calculation under different cutoffs prevents the most striking single estimate from becoming the whole story.

What can be reproduced?

The matched EDA models, metrics, bootstrap and descriptive reference curves run from bundled aggregate inputs. The core EDA input was rebuilt from the participant-supplied source workbook and matched all 3,344 rows. Regional date auditing requires the separately obtained case-level source, which is not distributed.

Map geometry and older contextual views remain archived source-derived extracts; a complete raw-file-to-every-screen pipeline is not claimed.

Reproducibility guide · Verification record · Model results CSV · Sensitivity results CSV