Caracterización de usuarios de servicios de salud en Bahía Blanca según efector de atención mediante aprendizaje automático interpretable
; González, Gisela Paula
; Geri, Milva 
Otros responsables
Fecha
2025
Editorial
Escuela de Perfeccionamiento en Investigación Operativa
Revista
Título del libro
Nombre del evento
Congreso Internacional XXXVIII Encuentro Nacional de Docentes en Investigación Operativa y XXXVI Escuela de Perfeccionamiento en Investigación Operativa
Fecha del evento
2025-10-22
Institución organizadora
Escuela de Perfeccionamiento en Investigación Operativa
Universidad Nacional del Sur
Universidad Nacional del Sur
ISSN
eISSN
ISBN
978-987-47251-6-5
Idioma
Español
Tipo del evento
Congreso
Clasificación temática
Palabras clave
Resumen
Garantizar un acceso oportuno, equitativo y de calidad a los servicios de salud es un objetivo central de los sistemas sanitarios (Anderson, 1973). No depende solo de contar con recursos, sino también de cómo se organiza y distribuye la demanda entre los distintos niveles de atención. En muchos contextos, las guardias hospitalarias se encuentran sobrecargadas, mientras que los centros de atención primaria presentan capacidad ociosa (Balakrishnan y Soderstrom, 2000). Este desajuste refleja no solo problemas de gestión, sino también la necesidad de fortalecer el rol de la atención primaria como puerta de entrada al sistema de salud. El objetivo del presente trabajo es determinar cuáles son las variables sanitarias y sociodemográficas que agrupan a los usuarios de determinadas clases de efectores de salud (‘unidad sanitaria’, ‘consultorio médico’, ‘internación hospitalaria’, ‘guardia hospitalaria’ y ‘otros tratamientos/estudios’) para luego predecir a qué efector concurrirá un individuo con determinadas características. Haciendo uso de 3609 observaciones y 11 variables explicativas provenientes de un relevamiento presencial, la metodología aplicada consistió en entrenar un modelo Random Forest (Breiman,2001) para clasificar el efector al que acudiría cada individuo entre 6 clases posibles (considerando adicionalmente a quienes no concurren a ningún efector como la clase ‘sin efector’). El conjunto de datos se particionó en 3000 observaciones de entrenamiento y el resto para testeo. A fin de garantizar una adecuada clasificación sobre el conjunto predictivo, se aplicó oversampling mediante la técnica RandomOverSampler para lograr un balance entre clases. Asimismo, dado el número limitado de observaciones, se adoptó un enfoque agnóstico para interpretar las predicciones del modelo. Con este propósito, se empleó SHAP (SHapley Additive exPlanations), lo que permitió cuantificar la importancia relativa de cada predictor en la toma de decisiones del algoritmo (Lundberg & Lee, 2017; Molnar, 2020). De los resultados obtenidos se desprende que el modelo de clasificación aplicado mostró un desempeño sólido para predecir a qué tipo de efector de salud concurrirá un individuo según sus características sanitarias y sociodemográficas, alcanzando una exactitud (accuracy) del 92%. Las clases ‘sin efector’ y ‘unidad sanitaria’ fueron clasificadas de manera perfecta, mientras que ‘consultorio médico’, ‘internación hospitalaria’, ‘guardia hospitalaria’ y ‘otros tratamientos/estudios’ presentaron algunas confusiones entre sí, aunque manteniendo altos valores de F1-score (0.88–0.90). El análisis de importancia de variables reveló que los factores más determinantes fueron el tipo de institución, el tipo de estudio y la distancia geográfica, seguidos por variables relacionadas con accesibilidad, edad y experiencia del usuario. Los resultados obtenidos permiten concluir que la metodología aplicada identifica con precisión los atributos clave que influyen en la elección de los efectores y constituye una base sólida para predecir, de manera individualizada, el lugar de atención sanitaria. De este modo, es posible anticipar a qué efector acudirá cada persona según sus características sociodemográficas y sanitarias, lo que contribuye a una gestión más eficiente de los recursos, previene la saturación de los servicios y favorece una planificación sanitaria más adecuada.
Ensuring timely, equitable, and quality access to health services is a central objective of health systems (Anderson, 1973). It depends not only on the availability of resources but also on how demand is organized and distributed across different levels of care. In many contexts, hospital emergency departments are overloaded, while primary care centers operate below capacity (Balakrishnan & Soderstrom, 2000). This imbalance reflects not only management issues but also the need to strengthen the role of primary care as the main gateway to the health system. The objective of this study was to determine which health and sociodemographic variables group users into specific categories of health care providers (“primary health unit,” “medical office,” “hospitalization,” “emergency department,” and “other treatments/examinations”), and subsequently predict the provider that an individual with certain characteristics would attend. Using 3,609 observations and 11 explanatory variables collected through an in- person survey, the methodology consisted of training a Random Forest model (Breiman, 2001) to classify the provider attended by each individual among six possible classes (including an additional class, “no provider,” for those not attending any). The dataset was split into 3,000 training observations and the remainder for testing. To ensure adequate classification on the predictive set, oversampling was applied using the RandomOverSampler technique to balance class distributions. Furthermore, given the limited sample size, an agnostic approach was adopted for model interpretation. For this purpose, SHAP (SHapley Additive exPlanations) was employed to quantify the relative importance of each predictor in the algorithm’s decision-making process (Lundberg & Lee, 2017; Molnar, 2020). The results showed that the classification model performed robustly in predicting the type of health provider an individual would attend based on health and sociodemographic characteristics, achieving an accuracy of 92%. The “no provider” and “primary health unit” classes were classified perfectly, while “medical office,” “hospitalization,” “emergency department,” and “other treatments/examinations” presented some overlap but maintained high F1-scores (0.88–0.90). The variable importance analysis indicated that the most influential factors were type of institution, type of study, and geographic distance, followed by accessibility, age, and user experience. In conclusion, the applied methodology accurately identified the key attributes influencing provider choice and established a solid foundation for individualized prediction of health care utilization. This approach enables anticipating the provider each person is likely to attend according to their health and sociodemographic characteristics, thereby supporting more efficient resource management, preventing service saturation, and promoting improved health care planning.
Ensuring timely, equitable, and quality access to health services is a central objective of health systems (Anderson, 1973). It depends not only on the availability of resources but also on how demand is organized and distributed across different levels of care. In many contexts, hospital emergency departments are overloaded, while primary care centers operate below capacity (Balakrishnan & Soderstrom, 2000). This imbalance reflects not only management issues but also the need to strengthen the role of primary care as the main gateway to the health system. The objective of this study was to determine which health and sociodemographic variables group users into specific categories of health care providers (“primary health unit,” “medical office,” “hospitalization,” “emergency department,” and “other treatments/examinations”), and subsequently predict the provider that an individual with certain characteristics would attend. Using 3,609 observations and 11 explanatory variables collected through an in- person survey, the methodology consisted of training a Random Forest model (Breiman, 2001) to classify the provider attended by each individual among six possible classes (including an additional class, “no provider,” for those not attending any). The dataset was split into 3,000 training observations and the remainder for testing. To ensure adequate classification on the predictive set, oversampling was applied using the RandomOverSampler technique to balance class distributions. Furthermore, given the limited sample size, an agnostic approach was adopted for model interpretation. For this purpose, SHAP (SHapley Additive exPlanations) was employed to quantify the relative importance of each predictor in the algorithm’s decision-making process (Lundberg & Lee, 2017; Molnar, 2020). The results showed that the classification model performed robustly in predicting the type of health provider an individual would attend based on health and sociodemographic characteristics, achieving an accuracy of 92%. The “no provider” and “primary health unit” classes were classified perfectly, while “medical office,” “hospitalization,” “emergency department,” and “other treatments/examinations” presented some overlap but maintained high F1-scores (0.88–0.90). The variable importance analysis indicated that the most influential factors were type of institution, type of study, and geographic distance, followed by accessibility, age, and user experience. In conclusion, the applied methodology accurately identified the key attributes influencing provider choice and established a solid foundation for individualized prediction of health care utilization. This approach enables anticipating the provider each person is likely to attend according to their health and sociodemographic characteristics, thereby supporting more efficient resource management, preventing service saturation, and promoting improved health care planning.
Colecciones
Aprobación
Revisión
Complementado por
Referenciado por
Licencia Creative Commons
Excepto donde se indique lo contrario, la licencia de este ítem se describe como by
