Artificial Neural Network Prediction of Daily Upper Respiratory Tract Infection Burden from Meteorological and Air Pollution Data: An Out-of-Time Test

Background/Objectives: Upper respiratory tract infections (URTIs) create seasonal emergency department demand. We assessed whether temporal representations of meteorological and air pollution variables improve artificial neural network (ANN) prediction of daily URTI presentations and generalize to a later period. Methods: Daily presentations at a tertiary hospital in Istanbul (October 2022–September 2025) were modelled using 30 combinations of three environmental domains and ten temporal representations. Expanding-window validation on 682 development days selected nine ANN finalists, evaluated on 365 subsequent out-of-time days. Prespecified permutation importance was evaluated; reference-model comparisons, uncertainty estimation, calibration and sensitivity analyses were post hoc. Results: The lowest ANN test mean absolute errors (MAEs) were 38.53 (Combined_D02), 38.78 (Weather_D01), and 39.75 (Pollution_S0) presentations/day; differences between leading models were small and imprecise. Test R2 ranged from 0.039 to 0.168, with negative bias in every model. On 45 high-demand days above the development-derived 90th percentile, all models underpredicted (observed mean 229.0; predicted means 95.6–111.6). A seven-day trailing mean (MAE 34.03), seasonal autoregressive integrated moving average (33.51), and autoregressive baseline (33.52) outperformed all ANNs; paired intervals for the two fitted baselines excluded zero. A calendar-only ANN achieved MAE 38.83, outperforming seven environmental models. Seasonally balanced splitting yielded lower MAEs (26.5–29.1), without consistent deterioration after training-size matching. Conclusions: No temporal representation consistently dominated. Environmental ANN models generalized poorly, underestimated peak demand and did not demonstrate gains over calendar predictors; simple outcome-history benchmarks were more accurate. They should not be used as stand-alone forecasting tools.

Authors

Institutions

Publication Details

Journal
Diagnostics
Published
2026-11-16
DOI
https://doi.org/10.3390/diagnostics16223800
Primary Topic
Climate Change and Health Impacts
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Artificial Neural Network Prediction of Daily Upper Respiratory Tract Infection Burden from Meteorological and Air Pollution Data: An Out-of-Time Test

Akif Bayyiğit, Sinan Karacabey, Tomris Albay, Mehmet Esat Ferhatlar et al.
1 links
Diagnostics
Climate Change and Health Impacts
article

Artificial Neural Network Prediction of Daily Upper Respiratory Tract Infection Burden from Meteorological and Air Pollution Data: An Out-of-Time Test

Akif Bayyiğit, Sinan Karacabey, Tomris Albay, Mehmet Esat Ferhatlar, Emre Kudu, ALPER TOSUN, Arzu Denizbaşı, Tufan Tükek, Merve ÇÖRTÜK, Mehmet Alptekin Acar, Yavuz Kutay Gökçe, Burak Hoş
article en

Abstract

Background/Objectives: Upper respiratory tract infections (URTIs) create seasonal emergency department demand. We assessed whether temporal representations of meteorological and air pollution variables improve artificial neural network (ANN) prediction of daily URTI presentations and generalize to a later period. Methods: Daily presentations at a tertiary hospital in Istanbul (October 2022–September 2025) were modelled using 30 combinations of three environmental domains and ten temporal representations. Expanding-window validation on 682 development days selected nine ANN finalists, evaluated on 365 subsequent out-of-time days. Prespecified permutation importance was evaluated; reference-model comparisons, uncertainty estimation, calibration and sensitivity analyses were post hoc. Results: The lowest ANN test mean absolute errors (MAEs) were 38.53 (Combined_D02), 38.78 (Weather_D01), and 39.75 (Pollution_S0) presentations/day; differences between leading models were small and imprecise. Test R2 ranged from 0.039 to 0.168, with negative bias in every model. On 45 high-demand days above the development-derived 90th percentile, all models underpredicted (observed mean 229.0; predicted means 95.6–111.6). A seven-day trailing mean (MAE 34.03), seasonal autoregressive integrated moving average (33.51), and autoregressive baseline (33.52) outperformed all ANNs; paired intervals for the two fitted baselines excluded zero. A calendar-only ANN achieved MAE 38.83, outperforming seven environmental models. Seasonally balanced splitting yielded lower MAEs (26.5–29.1), without consistent deterioration after training-size matching. Conclusions: No temporal representation consistently dominated. Environmental ANN models generalized poorly, underestimated peak demand and did not demonstrate gains over calendar predictors; simple outcome-history benchmarks were more accurate. They should not be used as stand-alone forecasting tools.

DiagnosticsVol. 16(22)
Marmara Üniversitesi Eğitim ve Araştırma Hastanesi (TR), Ankara Yıldırım Beyazıt University (TR), Liv Hospital (TR), Yildirim Beyazit University Yenimahalle Education and Research Hospital (TR), Istanbul Provincial Health Directorate (TR), Istanbul University (TR), Marmara University (TR)
Good health and well-being
Climate Change and Health Impacts
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon