APSR2026-1048

 
Late-Breaking
Submitted
Oral Presentation
AO32-007
 
 
https://storage.unitedwebnetwork.com/files/1285/1221262-46903-APSR2026_WS_1027_434926.pptx
Format: PPTX Only
 
Format: PDF Only
05. Respiratory Infections (non-tuberculous)
15

No more than 15 authors can be listed.
Please ensure the authors are listed in the right order.
If there are multiple affiliations, please use the "+" button next to "Select Country/Region" to add additional affiliation fields.

Wataru Satoh * Nagoya City University School of Data Science Nagoya Japan
Toshie Manabe - Nagoya City University Graduate School of Data Science Nagoya Japan
Anjarath Lorena Higuera-Iglesias - National Institute of Respiratory Diseases Epidemiology and Infectious Diseases Research Laboratory Mexico City Mexico
Thanh Van Do - Bach Mai Hospital Bach Mai Institute for Tropical Medicine Hanoi Vietnam
Luong Chinh Quoc - Bach Mai Hospital Neuro Intensive Care Department, National Institute of Neuroscience Hanoi Vietnam
Hiroki Manabe - Shitennoji University Faculty of Education Habikino Japan
Yoshihiro Kawade - Nagoya City University Graduate School of Medicine Nagoya Japan
Yukia Kawai - Nagoya City University School of Data Science Nagoya Japan
Rosario Fernández-Plata - National Institute of Respiratory Diseases Epidemiology and Infectious Diseases Research Laboratory Mexico City Mexico
David Martinez-Briseño - National Institute of Respiratory Diseases Epidemiology and Infectious Diseases Research Laboratory Mexico City Mexico
Manuel Castillejos-López - National Institute of Respiratory Diseases Epidemiology and Infectious Diseases Research Laboratory Mexico City Mexico
Son Ngoc Do - Bach Mai Hospital Center for Critical Care Medicine Hanoi Vietnam
Giap Van Vu - Bach Mai Hospital Bach Mai Respiratory Institute Hanoi Vietnam
Koichiro Kudo - Kawakita General Hospital Department of Respiratory Medicine Tokyo Japan
Co Xuan Dao - Bach Mai Hospital Center for Critical Care Medicine Hanoi Vietnam
Presenting Author Details
Wataru
Satoh
sthwtr@icloud.com
+818029322111
Under 30
Japan
Original Research
Development and Validation of a Machine Learning-Based COVID-19 Mortality Prediction Score in Independent Cohorts from Mexico and Vietnam
Maximum 25 words
Abstract Content
Accurate mortality risk stratification is crucial for allocating limited medical resources during infectious disease outbreaks. Existing scoring systems often rely on linear models that lack objective variable selection and struggle to capture complex nonlinearities. We aimed to develop a machine learning-based approach to derive accurate, clinically interpretable, and locally tailored COVID-19 mortality prediction scores using independent cohorts from Mexico and Vietnam, and to evaluate their cross-cohort applicability.
We retrospectively analyzed two independent cohorts: Mexico (3,046 inpatients, 35.0% fatality) and Vietnam (504 inpatients, 52.2% fatality). For each, an XGBoost model was trained on clinical and laboratory variables obtained at admission. SHAP values were utilized to extract key prognostic features. Integer-based scorecards were derived by scaling average SHAP values across feature quartiles, optimizing variable numbers via the elbow method. Cross-cohort applicability was evaluated by applying the Mexico score to the Vietnam cohort, and vice versa. AUCs were compared with established prognostic scores based on available variables.
The Mexico 5-variable score (ARDS, GCS, age, obesity, sepsis) achieved an internal AUC of 0.919, outperforming APACHE II (0.882) and SOFA (0.876). The Vietnam 5-variable score (ARDS, SpO2, GCS, age, serum ferritin) achieved an internal AUC of 0.907, outperforming the 4C Mortality Score (0.795), qSOFA (0.706), and CURB-65 (0.690). Notably, cross-cohort validations revealed a distinct performance decline: the Mexico score yielded an AUC of 0.671 in the Vietnam cohort, and the Vietnam score yielded 0.795 in the Mexico cohort. This decline was primarily driven by cohort-specific variables (e.g., sepsis vs. serum ferritin).
Integrating XGBoost with SHAP analysis enables the development of accurate, interpretable mortality scores tailored to local datasets. The notable decline in cross-cohort performance highlights the limitations of one-size-fits-all universal scores, strongly supporting local score derivation. This approach facilitates rapid risk stratification and optimizes resource allocation by adapting to setting-specific variables during outbreaks.
 
 
 
 
 
No
No
 
JPG or JPEG or PNG format ONLY
COVID-19, Prediction score, Machine Learning
maximum 3 keywords and no abbreviations, separated by commas
The authors declare no competing interests.
311