Guilherme Coelho
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Guilherme Coelho has a total of 22 co-authors. Here are the top co-authors by number of shared publications:
Clarimar José Coelho
6 shared publications- Danielle Satie Kassada
Professora Doutora da Área de Saúde Pública e Coletiva. Faculdade de Enfermagem da Universidade Estadual de Campinas – UNICAMP. Campinas/SP, Brasil
4 shared publications - Clarimar José Coelho
Scientific Computing Lab, Pontifical Catholic University of Goias,Goiania,Brazil
3 shared publications - Ana Carolina de Souza Peratelli
School of Medical Sciences, University of Campinas, Campinas 13083-888, Brazil
2 shared publications Charles M'POca Charles
2 shared publicationsDanielle Satie Kassada
2 shared publications
Latest Publications
- Open Access
- 15 November 2026
- Published by Sociedade Brasileira de Informática em Saúde in Journal of Health Informatics
- Vol. 18, 1684
Integrating machine learning demand forecasting and adaptive operational control in inpatient stroke care: an analysis with real hospital data
Objective: To compare capacity allocation policies for inpatient stroke care integrating Random Forest demand forecasting with machine-learning model predictive control (ML-MPC). Methods: Retrospective simulation study using 6,224 stroke admissions at a public neurological reference hospital in Southeastern Brazil (2006–2022). A daily series (2015–2022) was split into training (2015–2019), validation (2020) and test (2021–2022). Base capacity was derived from the 90th percentile of the observed daily census. Six policies were simulated over 727 days, with multi-step forecasts generated recursively from information available at each decision instant; uncertainty was assessed by moving-block bootstrap, re-executing all policies on resampled trajectories. Results: ML-MPC attained a mean simulated queue of 0.55 patients and the lowest objective-function value, 58% below the robust policy using 71% less extraordinary capacity (708 fewer units), and 46% below a naive fixed-capacity policy; 95% intervals excluded zero for blocks of 14, 30 and 60 days. Random Forest accuracy exceeded a historical-mean benchmark only marginally (MAE 0.994 vs 0.999). Conclusion: Within model conditions, how forecasts, uncertainty and capacity constraints entered sequential decisions mattered more than forecast accuracy. Outcomes are simulation estimates; no patient was exposed, the risk term is a non-validated relative index, and objective weights are judgment-based, not monetary.
- 15 November 2026
- Published by Sociedade Brasileira de Informática em Saúde in Journal of Health Informatics
- Vol. 18, 1684
Open AccessIntegrating machine learning demand forecasting and adaptive operational control in inpatient stroke care: an analysis with real hospital data
Objective: To compare capacity allocation policies for inpatient stroke care integrating Random Forest demand forecasting with machine-learning model predictive control (ML-MPC). Methods: Retrospective simulation study using 6,224 stroke admissions at a public neurological reference hospital in Southeastern Brazil (2006–2022). A daily series (2015–2022) was split into training (2015–2019), validation (2020) and test (2021–2022). Base capacity was derived from the 90th percentile of the observed daily census. Six policies were simulated over 727 days, with multi-step forecasts generated recursively from information available at each decision instant; uncertainty was assessed by moving-block bootstrap, re-executing all policies on resampled trajectories. Results: ML-MPC attained a mean simulated queue of 0.55 patients and the lowest objective-function value, 58% below the robust policy using 71% less extraordinary capacity (708 fewer units), and 46% below a naive fixed-capacity policy; 95% intervals excluded zero for blocks of 14, 30 and 60 days. Random Forest accuracy exceeded a historical-mean benchmark only marginally (MAE 0.994 vs 0.999). Conclusion: Within model conditions, how forecasts, uncertainty and capacity constraints entered sequential decisions mattered more than forecast accuracy. Outcomes are simulation estimates; no patient was exposed, the risk term is a non-validated relative index, and objective weights are judgment-based, not monetary.
- 5 November 2026
- Published by Society of Teachers of Family Medicine in Family Medicine
- Vol. 58 (10) , 806-807
Grammars of the Clinical Encounter: A Universal Language of Care
- 5 November 2026
- Published by Society of Teachers of Family Medicine in Family Medicine
- Vol. 58 (10) , 806-807
Grammars of the Clinical Encounter: A Universal Language of Care
- 31 October 2026
- Published by Elsevier in Chemometrics and Intelligent Laboratory Systems
- Vol. 280
Gate topology and reproducibility of Prior-Guided Spectral Gating across near-infrared and Raman spectroscopy
- 31 October 2026
- Published by Elsevier in Chemometrics and Intelligent Laboratory Systems
- Vol. 280
Gate topology and reproducibility of Prior-Guided Spectral Gating across near-infrared and Raman spectroscopy
- Open Access
- 23 October 2026
- Published by Elsevier in SSRN Electronic Journal
Prior Specification Bounds What Spectral Attention Can Learn
Learned per-wavelength weights are widely used to make chemometric neural networks interpretable. A common design passeseach weight through a sigmoid and initialises it from domain knowledge, expecting training to refine the prior against the spectra.We ask which weights the data can inform. The gradient reaching a weight is attenuated by s(1- s), so weights initialised near eitherend of the unit interval are effectively insulated from the data over realistic training budgets; where the gradient does arrive, spectralcollinearity constrains combinations of neighbouring weights rather than individual ones. We analyse a prior-guided spectral gatingmodel on near-infrared and Raman data, with synthetic verification against known ground truth. Over the bands given the maximumprior weight the data supply no measurable curvature — a contraction ratio of unity across the approximations, reference scales,optimisers and fitted solutions examined — while the curvature eigenvectors show the data determining regional combinationsof weights and leaving trades between neighbours free. Neither bound appears in predictive performance: with realistic spectralcorrelation, a prior contradicting the true relevance matches the correct one to within a pre-specified margin of 0.02 in held-out R2while placing 2% of the gate weight on the relevant regions against 72%; region-level permutation shows the network relying onthose regions regardless. Under the training conditions examined, prior specification therefore bounds what spectral attention canlearn, and predictive accuracy cannot validate the explanation. Normalising a prior by its maximum turns a flat top into a block ofinert parameters. Keeping the prior’s weights away from the extremes raises the agreement between the gate and an operationalreference for band relevance — the absolute coefficients of a calibrated partial least squares model, not a chemical ground truth —in every one of five partitions grouped by harvest population; the accompanying advantage in held-out accuracy does not survivethe same grouping, and we report it as specific to the partition on which it was measured.
- 23 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPrior Specification Bounds What Spectral Attention Can Learn
Learned per-wavelength weights are widely used to make chemometric neural networks interpretable. A common design passeseach weight through a sigmoid and initialises it from domain knowledge, expecting training to refine the prior against the spectra.We ask which weights the data can inform. The gradient reaching a weight is attenuated by s(1- s), so weights initialised near eitherend of the unit interval are effectively insulated from the data over realistic training budgets; where the gradient does arrive, spectralcollinearity constrains combinations of neighbouring weights rather than individual ones. We analyse a prior-guided spectral gatingmodel on near-infrared and Raman data, with synthetic verification against known ground truth. Over the bands given the maximumprior weight the data supply no measurable curvature — a contraction ratio of unity across the approximations, reference scales,optimisers and fitted solutions examined — while the curvature eigenvectors show the data determining regional combinationsof weights and leaving trades between neighbours free. Neither bound appears in predictive performance: with realistic spectralcorrelation, a prior contradicting the true relevance matches the correct one to within a pre-specified margin of 0.02 in held-out R2while placing 2% of the gate weight on the relevant regions against 72%; region-level permutation shows the network relying onthose regions regardless. Under the training conditions examined, prior specification therefore bounds what spectral attention canlearn, and predictive accuracy cannot validate the explanation. Normalising a prior by its maximum turns a flat top into a block ofinert parameters. Keeping the prior’s weights away from the extremes raises the agreement between the gate and an operationalreference for band relevance — the absolute coefficients of a calibrated partial least squares model, not a chemical ground truth —in every one of five partitions grouped by harvest population; the accompanying advantage in held-out accuracy does not survivethe same grouping, and we report it as specific to the partition on which it was measured.
- Open Access
- 22 September 2026
- Published by Springer Nature in Injury Epidemiology
- Vol. 13 (1) , 62
Epidemiological and etiological profile of spinal cord injury in Brazil: a systematic review and meta-analysis
Objectives: Spinal Cord Injury (SCI) severely impacts quality of life, yet consistent epidemiological data in Brazil are scarce. This study aimed to delineate the etiological profile of SCI in the Brazilian population to support public health strategies. Design: We conducted a systematic review and meta-analysis of proportions. Searches were performed in Medline, LILACS, and Embase. From the selected articles, 14 were included in the qualitative synthesis and 9 in the meta-analysis. The risk of bias was assessed using the JBI tool. Results: The meta-analysis with 8 studies revealed that the main etiologies of SCI were motor vehicle accidents (37.3%; 95% CI: 23.9–51.7), followed by falls (22.7%; 95% CI:16.2–29.9), violence including firearm injuries and physical assault (19.4%; 95% CI: 14.5–24.8), and diving into shallow water (10.9%; 95% CI: 5.1–18.3). Regarding the neurological level, the meta-analysis comprising 6 studies showed a pooled proportion of cervical injury of 37.4% (95% CI, 27.0-48.4). A male predominance was observed. We identified high heterogeneity (I² ranging from 76% to 95%) across all analyses, significant methodological weaknesses (e.g., hospital-based convenience samples), and evidence of publication bias for falls and violence. Conclusions: The etiological profile of SCI in Brazil reveals motor vehicle accidents as the leading cause, followed by falls and violence; and the most commonly affected neurological level is the cervical region. However, high heterogeneity and the risk of bias in primary studies demand a cautious interpretation of these findings. There is an urgent need for methodologically robust population studies to discover reliable data that can assist in effective public policies for the prevention of spinal cord injuries.
- 22 September 2026
- Published by Springer Nature in Injury Epidemiology
- Vol. 13 (1) , 62
Open AccessEpidemiological and etiological profile of spinal cord injury in Brazil: a systematic review and meta-analysis
Objectives: Spinal Cord Injury (SCI) severely impacts quality of life, yet consistent epidemiological data in Brazil are scarce. This study aimed to delineate the etiological profile of SCI in the Brazilian population to support public health strategies. Design: We conducted a systematic review and meta-analysis of proportions. Searches were performed in Medline, LILACS, and Embase. From the selected articles, 14 were included in the qualitative synthesis and 9 in the meta-analysis. The risk of bias was assessed using the JBI tool. Results: The meta-analysis with 8 studies revealed that the main etiologies of SCI were motor vehicle accidents (37.3%; 95% CI: 23.9–51.7), followed by falls (22.7%; 95% CI:16.2–29.9), violence including firearm injuries and physical assault (19.4%; 95% CI: 14.5–24.8), and diving into shallow water (10.9%; 95% CI: 5.1–18.3). Regarding the neurological level, the meta-analysis comprising 6 studies showed a pooled proportion of cervical injury of 37.4% (95% CI, 27.0-48.4). A male predominance was observed. We identified high heterogeneity (I² ranging from 76% to 95%) across all analyses, significant methodological weaknesses (e.g., hospital-based convenience samples), and evidence of publication bias for falls and violence. Conclusions: The etiological profile of SCI in Brazil reveals motor vehicle accidents as the leading cause, followed by falls and violence; and the most commonly affected neurological level is the cervical region. However, high heterogeneity and the risk of bias in primary studies demand a cautious interpretation of these findings. There is an urgent need for methodologically robust population studies to discover reliable data that can assist in effective public policies for the prevention of spinal cord injuries.
- Open Access
- 7 August 2026
- Published by Elsevier in SSRN Electronic Journal
Prior-Guided Spectral Gating Generalizes Across Spectroscopic Modalities: Evidence from Near-Infrared and Raman Spectroscopy
Prior-Guided Spectral Gating (PGSG) is a chemometrically-informed spec-tral attention mechanism, previously validated only on near-infrared (NIR)spectroscopy, where it outperforms partial least squares (PLS) at scale (n ≈1,448, R2 = 0.806 versus 0.568). Whether the mechanism generalizes tomodalities with markedly different spectral structure, such as Raman, hasnot been tested. We evaluate the unmodified architecture — an independentsigmoid gate per variable feeding a shallow multilayer perceptron, trainedend-to-end — on public NIR (n = 1,448, dry-matter content) and Raman(n = 6,472, glucose concentration) benchmarks, and formalize three descrip-tors of learned-gate topology (entropy, smoothness, sparsity) to characterizecross-modality structural differences. The model outperforms PLS in bothmodalities (∆R2 PLS = +0.244 NIR, +0.025 Raman); Raman gates are moresparse than NIR gates, and NIR gates are smoother than Raman gates oncesmoothness is normalized by physical band density, a correction we show tobe necessary and that reverses the naive raw-index comparison. A literature-informed prior does not improve accuracy over an uninformed initializationbut substantially increases gate reproducibility across training seeds (pairwiseJaccard similarity 0.90–0.94 versus 0.58–0.65). A finer, band-level test corre-lating local gate structure with literature-reported vibrational linewidths did not yield a robust correlation. These results indicate that PGSG generalizesacross spectroscopic modalities without architectural modification, that themagnitude of its advantage scales with how diluted the analyte signal alreadyis in a given modality, and that the principal value of a physically motivatedprior is interpretive stability rather than predictive gain.
- 7 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPrior-Guided Spectral Gating Generalizes Across Spectroscopic Modalities: Evidence from Near-Infrared and Raman Spectroscopy
Prior-Guided Spectral Gating (PGSG) is a chemometrically-informed spec-tral attention mechanism, previously validated only on near-infrared (NIR)spectroscopy, where it outperforms partial least squares (PLS) at scale (n ≈1,448, R2 = 0.806 versus 0.568). Whether the mechanism generalizes tomodalities with markedly different spectral structure, such as Raman, hasnot been tested. We evaluate the unmodified architecture — an independentsigmoid gate per variable feeding a shallow multilayer perceptron, trainedend-to-end — on public NIR (n = 1,448, dry-matter content) and Raman(n = 6,472, glucose concentration) benchmarks, and formalize three descrip-tors of learned-gate topology (entropy, smoothness, sparsity) to characterizecross-modality structural differences. The model outperforms PLS in bothmodalities (∆R2 PLS = +0.244 NIR, +0.025 Raman); Raman gates are moresparse than NIR gates, and NIR gates are smoother than Raman gates oncesmoothness is normalized by physical band density, a correction we show tobe necessary and that reverses the naive raw-index comparison. A literature-informed prior does not improve accuracy over an uninformed initializationbut substantially increases gate reproducibility across training seeds (pairwiseJaccard similarity 0.90–0.94 versus 0.58–0.65). A finer, band-level test corre-lating local gate structure with literature-reported vibrational linewidths did not yield a robust correlation. These results indicate that PGSG generalizesacross spectroscopic modalities without architectural modification, that themagnitude of its advantage scales with how diluted the analyte signal alreadyis in a given modality, and that the principal value of a physically motivatedprior is interpretive stability rather than predictive gain.
- Open Access
- 7 July 2026
- Published by Elsevier in SSRN Electronic Journal
Prior-Guided Spectral Gating with End-to-EndOptimization for NIR Spectroscopy: ScalableArchitecture for Interpretable Regression inSmall-Sample Regimes
Prior-Guided Spectral Gating (PGSG) incorporates domain knowledge intonear-infrared (NIR) regression through a learnable gate vector initializedfrom a spectroscopic prior. We identify two fundamental limitations of theoriginal formulation that prevent its application to high-dimensional spectra:a softmax gradient vanishing problem (at p = 281 bands, the softmax Jaco-bian diagonal is 58× smaller than sigmoid, rendering the gate untrainable),and a circular prior pitfall (priors derived from training data cause the gateto reproduce the prior without adaptation). We propose PGSGv2, which re-places softmax with independent per-band sigmoid gates and a PLS regres-sor with an end-to-end MLP, enabling backpropagation through the jointgate–regressor objective. Applied to mango dry matter content (DMC%)prediction using the Mango DMC v3 dataset (n = 1,448, p = 281 effec-tive bands, within-season protocol) with a literature-based prior, PGSGv2achieves R2 = 0.806 at n = 1,159, outperforming PLS (R2 = 0.568, +42%)and MLP (R2 = 0.699, +15%). For n ≤ 60, PGSGv2 is the only methodwith positive R2 , demonstrating the value of prior-guided initialization indata-scarce regimes. Learned gates are chemically coherent (ρ(g, s) ≥ 0.998)and highly reproducible across seeds (Jaccard ≥ 0.945). PGSGv2 providesa scalable and interpretable architecture for NIR calibration when labeledsamples are limited.
- 7 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPrior-Guided Spectral Gating with End-to-EndOptimization for NIR Spectroscopy: ScalableArchitecture for Interpretable Regression inSmall-Sample Regimes
Prior-Guided Spectral Gating (PGSG) incorporates domain knowledge intonear-infrared (NIR) regression through a learnable gate vector initializedfrom a spectroscopic prior. We identify two fundamental limitations of theoriginal formulation that prevent its application to high-dimensional spectra:a softmax gradient vanishing problem (at p = 281 bands, the softmax Jaco-bian diagonal is 58× smaller than sigmoid, rendering the gate untrainable),and a circular prior pitfall (priors derived from training data cause the gateto reproduce the prior without adaptation). We propose PGSGv2, which re-places softmax with independent per-band sigmoid gates and a PLS regres-sor with an end-to-end MLP, enabling backpropagation through the jointgate–regressor objective. Applied to mango dry matter content (DMC%)prediction using the Mango DMC v3 dataset (n = 1,448, p = 281 effec-tive bands, within-season protocol) with a literature-based prior, PGSGv2achieves R2 = 0.806 at n = 1,159, outperforming PLS (R2 = 0.568, +42%)and MLP (R2 = 0.699, +15%). For n ≤ 60, PGSGv2 is the only methodwith positive R2 , demonstrating the value of prior-guided initialization indata-scarce regimes. Learned gates are chemically coherent (ρ(g, s) ≥ 0.998)and highly reproducible across seeds (Jaccard ≥ 0.945). PGSGv2 providesa scalable and interpretable architecture for NIR calibration when labeledsamples are limited.
- Open Access
- 8 June 2026
- Published by Society of Teachers of Family Medicine in Family Medicine
- Vol. 58 (6) , 452-453
Between Languages
- 8 June 2026
- Published by Society of Teachers of Family Medicine in Family Medicine
- Vol. 58 (6) , 452-453
Open AccessBetween Languages
- Open Access
- 29 May 2026
- Published by Elsevier in SSRN Electronic Journal
Prior-Guided Spectral Gating for Interpretable Deep Learning in Near-Infrared Spectroscopy
Deep learning models for near-infrared (NIR) spectroscopy often functionas black boxes, limiting adoption in chemometrics where interpretability isparamount. We propose a lightweight neural architecture with a prior-guidedspectral gating layer that combines the predictive power of deep learning withchemometric interpretability. Unlike standard attention mechanisms, ourgates are initialized from established variable importance measures (ANOVAF-statistic, VIP scores, Random Forest importance) and regularized via KLdivergence to maintain statistical grounding throughout training. We vali-date the framework on two benchmark NIR datasets from distinct domains:Tecator (food quality, n = 215, fat prediction) and Gasoline (petrochemical,n = 60, octane rating). Cross-validation results demonstrate competitiveperformance with PLS (R2 = 0.86 ± 0.04 vs 0.92 on Tecator; R2 = 0.93 vs0.94 on Gasoline) while providing explicit band importance maps. Learnedgates identify chemically relevant spectral regions, with the top selected bandfor Tecator at 931 nm corresponding to known C–H absorption for fat con-tent. The moderate correlation between learned gates and chemometricpriors (0.35–0.51) confirms that the model refines domain knowledge withdata-driven learning rather than merely reproducing initialization. Our ap-proach offers a practical compromise between black-box deep learning andtraditional linear methods, particularly valuable for small-sample chemomet-ric applications where interpretability and domain knowledge integration areessential.
- 29 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPrior-Guided Spectral Gating for Interpretable Deep Learning in Near-Infrared Spectroscopy
Deep learning models for near-infrared (NIR) spectroscopy often functionas black boxes, limiting adoption in chemometrics where interpretability isparamount. We propose a lightweight neural architecture with a prior-guidedspectral gating layer that combines the predictive power of deep learning withchemometric interpretability. Unlike standard attention mechanisms, ourgates are initialized from established variable importance measures (ANOVAF-statistic, VIP scores, Random Forest importance) and regularized via KLdivergence to maintain statistical grounding throughout training. We vali-date the framework on two benchmark NIR datasets from distinct domains:Tecator (food quality, n = 215, fat prediction) and Gasoline (petrochemical,n = 60, octane rating). Cross-validation results demonstrate competitiveperformance with PLS (R2 = 0.86 ± 0.04 vs 0.92 on Tecator; R2 = 0.93 vs0.94 on Gasoline) while providing explicit band importance maps. Learnedgates identify chemically relevant spectral regions, with the top selected bandfor Tecator at 931 nm corresponding to known C–H absorption for fat con-tent. The moderate correlation between learned gates and chemometricpriors (0.35–0.51) confirms that the model refines domain knowledge withdata-driven learning rather than merely reproducing initialization. Our ap-proach offers a practical compromise between black-box deep learning andtraditional linear methods, particularly valuable for small-sample chemomet-ric applications where interpretability and domain knowledge integration areessential.
- Open Access
- 18 May 2026
- Published by Public Library of Science (PLoS) in PLOS Climate
- Vol. 5 (5) , e0000801
The impact of extreme temperatures on respiratory mortality in Brazil: Evaluating regional adaptations to different thermal environments
Ambient temperature is associated with respiratory mortality, but evidence from tropical middle-income countries remains limited. We conducted a nationwide ecological time-series study to quantify associations between temperature and respiratory mortality across Brazil's diverse climates and to estimate the attributable burden. We analysed 1,087,094 respiratory deaths (ICD-10 J00–J99) from 646 municipalities (population ≥50,000) spanning all five Brazilian macroregions from 2010 to 2020. Daily temperature-mortality associations were estimated using distributed lag non-linear models (DLNM) with quasi-Poisson regression, adjusting for seasonality, long-term trends, and day of week. City-specific estimates were pooled using random-effects meta-analysis, and attributable fractions were calculated using the minimum mortality temperature (MMT) as reference. The pooled exposure-response curve showed a J-shaped relationship between temperature and respiratory mortality. The national MMT was 22.4°C, with low heterogeneity across municipalities (I² = 17.3%). Cumulative relative risks over lags 0–21 days were 1.29 (95% CI: 1.23–1.35) at the 1st temperature percentile (12.1°C) and 1.43 (95% CI: 1.36–1.50) at the 99th percentile (30.3°C). Overall, 6.08% of respiratory deaths were attributable to non-optimal temperatures, corresponding to 66,079 excess deaths during the study period. Regional patterns varied markedly: the North exhibited heat-dominant vulnerability (attributable fraction 12.5%), whereas the South showed cold-dominant effects (7.5%). Heat exposure accounted for a larger share of the national burden (4.27%) than cold (1.81%), reflecting Brazil's predominantly tropical climate. Temperature extremes substantially increase respiratory mortality across Brazil, with distinct regional vulnerability profiles. These findings support the implementation of region-specific early warning systems addressing both thermal extremes and inform climate adaptation strategies for tropical settings.
- 18 May 2026
- Published by Public Library of Science (PLoS) in PLOS Climate
- Vol. 5 (5) , e0000801
Open AccessThe impact of extreme temperatures on respiratory mortality in Brazil: Evaluating regional adaptations to different thermal environments
Ambient temperature is associated with respiratory mortality, but evidence from tropical middle-income countries remains limited. We conducted a nationwide ecological time-series study to quantify associations between temperature and respiratory mortality across Brazil's diverse climates and to estimate the attributable burden. We analysed 1,087,094 respiratory deaths (ICD-10 J00–J99) from 646 municipalities (population ≥50,000) spanning all five Brazilian macroregions from 2010 to 2020. Daily temperature-mortality associations were estimated using distributed lag non-linear models (DLNM) with quasi-Poisson regression, adjusting for seasonality, long-term trends, and day of week. City-specific estimates were pooled using random-effects meta-analysis, and attributable fractions were calculated using the minimum mortality temperature (MMT) as reference. The pooled exposure-response curve showed a J-shaped relationship between temperature and respiratory mortality. The national MMT was 22.4°C, with low heterogeneity across municipalities (I² = 17.3%). Cumulative relative risks over lags 0–21 days were 1.29 (95% CI: 1.23–1.35) at the 1st temperature percentile (12.1°C) and 1.43 (95% CI: 1.36–1.50) at the 99th percentile (30.3°C). Overall, 6.08% of respiratory deaths were attributable to non-optimal temperatures, corresponding to 66,079 excess deaths during the study period. Regional patterns varied markedly: the North exhibited heat-dominant vulnerability (attributable fraction 12.5%), whereas the South showed cold-dominant effects (7.5%). Heat exposure accounted for a larger share of the national burden (4.27%) than cold (1.81%), reflecting Brazil's predominantly tropical climate. Temperature extremes substantially increase respiratory mortality across Brazil, with distinct regional vulnerability profiles. These findings support the implementation of region-specific early warning systems addressing both thermal extremes and inform climate adaptation strategies for tropical settings.