An exploratory platform combining official Peruvian epidemiological surveillance, environmental observations and territorial exposure information. It tests whether machine learning can improve early warning rather than claiming it already does.
An exploratory research prototype, not an operational warning system.
Explore the weekly EDA benchmark in 04, inspect the regional source audit in 06, and read the uncertainty and sensitivity analysis in 08. Models are run offline in Python; this interface displays their saved outputs.
Research note (PDF) · One-page project brief · Verification record
Trends from the official MINSA historical files. Dengue and leptospirosis show notifications; EDA is a count of reported episodes, not individual people.
*2024 EDA 5+: not comparable. The MINSA export contains zero entries after week 15, inconsistent with a CDC Peru dashboard export reporting 57,363 in 2024 and the DIRESA Piura SE52 report of 56,726. See the source-audit table below.
The CDC Peru dashboard export provides a second official series for the same displayed indicator. It differs slightly from the older national open-data extract in 2019–2023 and dramatically in 2024. The original files are preserved; no missing weekly values have been imputed.
| Year | Older MINSA CSV | CDC dashboard Excel | Difference |
|---|---|---|---|
| 2019 | 41,710 | 41,935 | +225 |
| 2020 | 20,085 | 20,271 | +186 |
| 2021 | 24,694 | 25,294 | +600 |
| 2022 | 32,699 | 32,894 | +195 |
| 2023 | 54,311 | 54,509 | +198 |
| 2024 | 15,451 | 57,363 | +41,912 |
Root symptom: in Piura, the older national CSV records zero EDA episodes for over-age-5 from weeks 16–52 of 2024. The CDC dashboard shows nonzero weekly activity throughout that period. Zero in the CSV therefore indicates missing/incomplete surveillance content, not no disease.
Methodological boundary: the 57,363 figure is a validated annual dashboard total, not an imported district-week series. It is deliberately not inserted into the weekly machine-learning dataset; weekly 2024 EDA over-age-5 remain excluded pending a trustworthy complete export. Small differences in 2019–2023 may reflect surveillance revisions; their causes have not been determined.
The October 2026 CDC Perú annual indicator download confirms 33,216 episodes and 307 hospitalizations in 2024, compared with 33,059 episodes and 297 hospitalizations in the older MINSA CSV. Both age groups are reviewed below; these are different publication snapshots, not counts to sum across sources.
| Year | Older MINSA CSV, under 5 | CDC dashboard Excel, under 5 | Difference |
|---|---|---|---|
| 2019 | 27,181 | 27,386 | +205 |
| 2020 | 14,718 | 14,846 | +128 |
| 2021 | 19,424 | 19,603 | +179 |
| 2022 | 24,370 | 24,519 | +149 |
| 2023 | 30,604 | 30,778 | +174 |
| 2024 | 33,059 | 33,216 | +157 |
Important distinction: all 52 weeks in the 2024 MINSA under-5 series have nonzero episode counts, unlike the incomplete older age-5+ series. Under-5 annual differences are below 1% across 2019–2024, but the reason for the differences is not yet established. The annual chart above still presents the older MINSA series, and its 2024 under-5 value of 33,059 has not been silently overwritten with the CDC annual total.
2024 admissions alert for research: the CDC dashboard records 307 under-5 hospitalizations (2023: 77). This is a substantial surveillance change that warrants investigation of referral, admission, and reporting practices; it is not proof of increased disease severity or an El Niño effect.
Evidence: CDC Perú – EDA situation dashboard · National MINSA historical open dataset · annual CDC Excel downloaded October 2026. See CDC_EDA_Piura_2019_2024_source_comparison.csv and audit_eda_piura_both_age_groups_2019_2024.csv.
Original MINSA open-data extract updated 21 October 2026. The 2000–2024 national and Piura annual counts match the older file; 2025 is newly available. Notification counts are not unique patients or rates.
Data source: MINSA dengue open-data catalog. Supporting reproducible extracts: dengue_piura_departamento_semanal_2016_2025.csv and data/dengue_piura_provincia_anual_2016_2025.csv.
Four-week-ahead dengue forecasting experiment for Piura. Training: 2018–2022. Held-out evaluation: 2023–2024.
Population-weighted mean of weekly precipitation across 65 districts; this is not the total rain volume of the region.
CENEPRED 2026 physical-risk scenarios: flooding and mass movements. These levels are not forecasts of disease outbreaks.
Official CDC Peru 2018–2025 EDA episodes with observed climate, presented at province level. This panel uses all-age EDA; age-stratified data are available as annual summaries only.
Rainfall cannot by itself prove that a change in EDA was caused by El Niño. Missing climatic observations are not treated as zero.
Out-of-time retrospective test, four epidemiological weeks ahead. Same province, target week and evaluation sample for both methods.
Controlled retrospective ablation: identical 992 province-weeks (2023-2025), same four-week-ahead prediction target, Random Forest configuration and training years (2019-2022). Only the climate predictors change. The chart and all release documents now use this same matched experiment (climate Random Forest MAE 50.29).
| Method | MAE (episodes) | Interpretation |
|---|---|---|
| Last observed week | 45.09 | Best of tested methods |
| Four-week trailing average | 46.67 | Simple baseline |
| Random Forest + observed climate | 50.29 | Lower observed error within this RF comparison |
| Random Forest, EDA-only | 52.76 | Identical algorithm, without weather |
| Previous-year same week | 71.95 | Seasonal naive baseline |
Reproduce with python research/eda_climate_ablation.py. Download detailed metric comparison. Methodological references: US CDC - Appraising models; WHO EWARS.
The fixed v0.8 out-of-time climate Random Forest has different error profiles by province. Positive % means the climate forest reduces MAE relative to simple persistence, not that climate alone caused the difference. All 2023–2025 exploratory observations and algorithm parameters remain unchanged.
| Piura province | Evaluations | Persistence MAE | RF + climate MAE | Change vs persistence |
|---|---|---|---|---|
| Ayabaca | 124 | 19.2 | 17.6 | +8.2% |
| Huancabamba | 124 | 16.6 | 16.0 | +4.2% |
| Morropon | 124 | 31.2 | 28.3 | +9.3% |
| Paita | 124 | 28.8 | 25.6 | +11.3% |
| Piura | 124 | 138.1 | 180.4 | -30.7% |
| Sechura | 124 | 11.7 | 12.0 | -2.7% |
| Sullana | 124 | 73.7 | 82.6 | -12.1% |
| Talara | 124 | 41.4 | 39.9 | +3.8% |
Recalculate from the existing matched ablation predictions. Download province-level error metrics.
El Nino is historical context for this project. No 2027 climate or epidemic forecast is generated or validated here. Consult official authorities for current forecasts.
What NinoSalud can support: a transparent review of historical surveillance, climatic anomalies, source discrepancies and district exposure classifications. It cannot generate certified outbreak warnings or predict 2027 case counts.
Official sources for current context: ENFEN announcements · MINSA CDC Peru · IMARPE coastal El Nino index (ICEN). Official outlooks are time-sensitive.
Exploratory weekly comparison, not an epidemic detector. For each province and target week in 2023–2025, the historical reference is computed from 2018–2022 only, using observations from the same week plus/minus two weeks. Week 53 is mapped to week 52 in the historical reference. The 90th percentile is a descriptive threshold, not an estimated probability.
Transparent Python implementation in research/eda_reference_signals.py; download all province-week comparisons. Further database opportunities and limitations in docs/NORTHERN_PERU_SOURCE_RESEARCH.md.
A new independent audit of the GORE Piura regional dengue dataset (2023, 2025 and partial 2026), with source dictionary cross-checks. This measures the interval between the two recorded dates, not confirmed disease incidence, central reporting lag, treatment waiting time or forecast effectiveness.
Compare patterns cautiously: case definitions, rural access, incomplete reporting, epidemic timing, and date quality can all affect delays. These comparisons are descriptive, not causal.
| Field | Definition in supplied official dictionary | Responsible interpretation |
|---|---|---|
| FECHA_INICIO_SINTOMAS | Patient's initial symptom date | Self-reported or recorded onset date |
| FECHA_NOTIFICACION | Date patient attended health center | Recorded attendance proxy, not proven system notification timestamp |
| FECHA_INGRESO | Date entered into information system | Chronology inconsistent in many rows; no reliable delay conclusion |
| ANO | Year case was recorded | Not always identical to calendar year of attendance |
We found that the provided dictionary also describes UBIGEO using wording about a hydrological station and gives the notification-week field length as one character. These appear to be metadata-template errors and require source validation.
Method and aggregate-only exports in docs/REGIONAL_DENGUE_2026_AUDIT.md and data/regional_dengue_*.csv. Download province summaries. Original case-level rows are not bundled; source: GORE Piura open data.
Cristian Daniel Medina Gomez · Independent project informed by undergraduate studies in Economics and Management at Bocconi. No institutional supervision or endorsement is implied.
Project development: the project owner selected the Peru-focused direction, collected public source materials, requested iterative checks, reviewed the presentation and tested the interface.
AI-assisted research and implementation: generative AI provided substantial assistance with programming, analytical implementation, documentation and packaging. Methods, limitations and reproducible scripts are provided so that the outputs can be examined rather than accepted on trust.
The full contribution statement is included in docs/CONTRIBUTIONS.md.
Dengue model: historical Piura-wide notifications, trained on 2018–2022 and tested on 2023–2024; displayed in tab 02. EDA model: provincial all-age episode counts, trained on 2019–2022 and tested on 2023–2025; displayed in tab 04. Both are exploratory and neither has demonstrated generalizable operational early warning. The 2026 exploratory ablation compares a matched EDA-only Random Forest against the same algorithm with observed weather features (tab 04).
EDA 2025 includes missing observed climate weeks. The EDA model evaluation excludes observations requiring those inputs, so its test metrics do not cover every week.
Official sources below would support a next research phase. A new regionally scoped dengue dataset has now been audited (tab 06); other catalogued sources remain unintegrated.
| Source | Potential question | Status |
|---|---|---|
| DIRESA Piura · dengue notifications | Symptom onset vs recorded attendance interval | Audited; tab 06 |
| GORE Piura · establishment EDA | Facility distribution of weekly episodes | Catalogue only; example 2024 file |
| SENAMHI · Piura station archive | Weather station coverage and quality | Catalogue only |
| CHIRPS v3 · rainfall grids | Independent rainfall sensitivity check | Catalogue only |
| CDC Perú · October 2026 dengue dashboard | Compare current surveillance across Piura, Tumbes, Lambayeque, La Libertad | Catalogue only; not a historical predictor |
| SENAMHI · Río Piura hydrology | River gauge state and possible flood exposure | Catalogue only |
Full 11-source inventory and matching warnings in docs/NORTHERN_PERU_SOURCE_RESEARCH.md.