WCP 2026 / STOCKHOLM / #3134
Can weekly surveillance data help us look ahead? Explore forecasts of depression, Alzheimer’s and Parkinson’s notifications in Mexico—and why local evaluation matters.
E-Poster 3134 · AS17 Digital Psychiatry
Guided by compassion, grounded in science: psychiatry for our time.wcp-congress.com ↗
26th World Congress of Psychiatry · 23–26 September 2026
AS17 Digital Psychiatry · Research, not a clinical decision tool.
01 / THE POSTER
Read, zoom in, or take a copy with you. Both PDFs contain the same study in their respective languages.
Poster edition: 8 September 2026. Frozen conference results, not the latest weekly report. PDF downloads remain available without JavaScript.
Population-level prediction of depression, Alzheimer’s disease, and Parkinson’s disease incidence using artificial intelligence: implications for health service planning
The registered title is preserved. The measured outcome is newly reported cases—not clinical incidence, measured care demand or individual risk.
02 / EXPLORE THE EVIDENCE
Compare weekly notifications with model estimates. Switch conditions and evaluation windows to see where performance changes.
Loading the frozen study data…
Before the last compared observation: retrospective model fit. Shaded future: forecasts, not observed outcomes. Model interval coverage has not been evaluated. Dates follow the poster’s bulletin calendar.
| Condition | sMAPE · 25 | sMAPE · 8 | MAE · 25 | MAE · 8 |
|---|---|---|---|---|
| Depression | 8.7% | 5.0% | 260.2 | 153.8 |
| Parkinson’s | 10.6% | 6.7% | 16.5 | 10.3 |
| Alzheimer’s | 14.4% | 17.8% | 6.2 | 7.4 |
sMAPE — symmetric mean absolute percentage error. Weekly relative error, averaged across evaluated weeks. Lower is better; 0% is an exact match. It is not “percent accuracy”.
MAE — mean absolute error. The average weekly difference in cases, ignoring its sign. It makes the size of the error concrete. Lower is better.
03 / LOOK BEYOND THE TOTAL
Depression, epidemiological weeks 7–31. Explore 32 states; the national model is independent, not the sum of the state forecasts.
Descriptive comparison of selected models, not uniform prospective out-of-sample validation. With few reported cases, small absolute errors can become large relative differences. This is a possible mechanism, not a demonstrated explanation for any particular state.
A DIFFERENT LENS / SEX
Depression · cumulative 2026 notifications through epidemiological week 31.
Different window and variable. These totals are not directly comparable with the national forecast callout.
Reported cases—not forecasts, individual risk or predictive performance by sex. This cumulative window differs from the 25-week evaluation. Even when the weeks are aligned, sex-specific cumulative increments do not reconcile with the weekly-case column; the cause has not been established.
04 / HOW THE STUDY WORKS
A retrospective study of Mexico’s weekly epidemiological notifications, 2014–2026. Each time series follows a condition, territory and sex stratum to reflect patterns a national total may miss.
A statistical forecasting method that represents trends and recurring cycles.
A machine-learning method: a recurrent neural network learns patterns across related time series.
Combines predictions from several methods to bring different forecasting perspectives together.
Uses a learned combination of model predictions rather than relying on a single approach.
A model is selected for each series. National results here use DeepAR for depression and Prophet for Parkinson’s and Alzheimer’s. We do not claim that every historical selection is the universally best model.
333 designed series = 3 conditions × 37 areas × 3 strata. Areas: 32 states, the national total and 4 groups of states. Strata: men, women and both sexes combined. Forecasts were retained for 297 series; 36 regional series were not evaluated because geographic names differed between files. Results here cover three national conditions and depression in 32 states, both sexes combined.
The 25-week window (weeks 7–31 of 2026) describes agreement with notifications. For weeks 24–31, evaluated predictions match the file dated 5 June 2026 within the documented rounding tolerance. The file predates those weeks; it does not establish the full training chronology or a cryptographically certified timestamp. The evaluation is not fully prospective.
Notifications are not clinical incidence or measured care demand. Access, reporting and underreporting may affect counts. Age and transfer outside Mexico were not evaluated. Clinical benefit, staffing impact, savings and budget-allocation effects were not measured.
The frozen 52-week forecast runs from 9 February 2026 to 1 February 2027. It is not a new forecast starting when you visit this page. Historical model fits are not presented as past prospective forecasts.
Sex-specific cumulative columns and the weekly general-case column do not reconcile exactly over the evaluation window. Their difference has not been assigned to a cause or artificially reconciled. No forecast is allocated by the proportions in the descriptive sex bar.
LET’S TALK AT WCP
Discuss local validation, the forecasting approach, or what a service-planning evaluation should measure.
rebull@exatec.tec.mx ↗