Intelligent monitoring of diabetes mellitus by means of mobile and wearable devices
- Autores: Ciro Rodríguez León
- Directores de la Tesis: Oresti Baños Legrán (codir. tes.), Claudia Villalonga Palliser (codir. tes.)
- Lectura: En la Universidad de Granada ( España ) en 2024
- Idioma: inglés
- ISBN: 9788411956932
- Número de páginas: 279
- Tribunal Calificador de la Tesis: Héctor Pomares Cintas (presid.), Javier Medina Quero (secret.), Jesús Peral Cortés (voc.), David Gil Méndez (voc.), Macarena Espinilla Estévez (voc.)
- Programa de doctorado: Programa de Doctorado en Tecnologías de la Información y la Comunicación por la Universidad de Granada
- Materias:
- Enlaces
- Tesis en acceso abierto en: DIGIBUG
- Resumen
- español
La diabetes tipo 1 se caracteriza fundamentalmente por un severo problema de secreción de insulina. Los pacientes con diabetes tipo 1 se ven obligados a utilizar una fuente externa de insulina para equilibrar sus niveles de glucemia en sangre, que es una tarea muy compleja debido a la alta variabilidad de patrones en diferentes pacientes, influenciada por factores como el estilo de vida, metabolismo, o sexo, entre otros. Es por ello que es una práctica común que muchos pacientes con diabetes tipo 1 usen dispositivos de monitorización continua de glucemia para gestionar mejor la enfermedad, ya que les permite conocer los niveles de glucemia en todo momento. Sin embargo, contar con una aplicación que prediga niveles de glucemia sería muy útil, ayudando a los pacientes a tomar medidas preventivas y evitar complicaciones por niveles fuera de rango. La predicción de estos niveles de glucemia en sangre es una tarea compleja. Por un lado, los algoritmos de predicción generales, que utilizan datos de mediciones continuas de glucemia de múltiples pacientes para entrenar y evaluar, no logran capturar por completo las diferencias individuales en estos patrones. Por otro lado, la creación de modelos de predicción personalizados para cada paciente resulta poco práctica, ya que requiere una gran cantidad de datos de mediciones continuas de glucemia, lo que implica que los pacientes nuevos tendrían que esperar mucho para obtener predicciones precisas. Además, existen muy pocos conjuntos de datos de acceso abierto con suficiente cantidad de pacientes y longitudinalidad de mediciones de glucemia continua como para hacer experimentaciones robustas.
En este contexto, el objetivo de esta tesis es investigar técnicas inteligentes para apoyar la gestión de la diabetes mellitus tipo 1 en entornos del mundo real aprovechando los datos de mediciones continuas de glucemia recogidos de dispositivos móviles y vestibles. Para cumplir este objetivo principal se plantean los siguientes objetivos específicos: (1) crear un amplio conjunto de datos longitudinales de larga duración y de libre acceso consistente en mediciones de los niveles de glucemia en sangre recogidas mediante sensores de monitorización continua de la glucemia para desarrollar nuevos modelos inteligentes, (2) comprender la capacidad de los métodos de evaluación existentes para estimar el rendimiento de los modelos de predicción de los niveles de glucemia en sangre en entornos del mundo real, (3) identificar grupos de pacientes con características comunes a partir de las mediciones longitudinales de los niveles de glucemia en sangre de los pacientes, (4) evaluar el rendimiento de los modelos de predicción del nivel de glucemia en sangre basados en grupos en entornos del mundo real. Para dicha evaluación, se desarrollarán y compararán tres tipos de modelos de predicción: generales, basados en grupos y personalizados, utilizando métricas de evaluación clínicas y no clínicas. De este modo se determina la viabilidad de la propuesta realizada para su uso en predicción de niveles de glucemia en sangre en entornos realistas.
Para abordar el problema, primeramente, se recopila un conjunto de datos con suficiente cantidad de mediciones continuas de glucemia y un número adecuado de pacientes, garantizando así una base sólida para realizar la experimentación planteada. A continuación, se desarrollan modelos de predicción de glucemia generales, donde el entrenamiento y evaluación se realizan utilizando datos de mediciones de glucemia continua de todos los pacientes del conjunto de datos. Esto permite identificar las estrategias de evaluación más apropiadas para este tipo de modelos de predicción. Una vez comprendidas estas estrategias, se procede a la creación de grupos de pacientes mediante el uso de algoritmos de aprendizaje no supervisado de diversa naturaleza. Dado que se prueban múltiples configuraciones de tamaños de grupos, versiones de los datos y algoritmos, el conjunto de resultados de agrupamiento es amplio. Por ello, se seleccionan los resultados más prometedores utilizando una métrica de evaluación interna del agrupamiento, complementadas con un procedimiento diseñado en esta tesis, que posteriormente son revisados por expertos clínicos para identificar la solución más adecuada. Con los grupos definidos, se replica la experimentación realizada con los modelos de predicción generales para evaluar el rendimiento comparativo entre estos y los modelos de predicción basados en grupos. Además, se seleccionan pacientes dentro de cada grupo para desarrollar modelos de predicción personalizados, lo que permite realizar una comparación adicional con los modelos de predicción basados en grupos y así determinar si estos últimos representan una solución intermedia viable entre los modelos de predicción generales y los personalizados. A continuación se detalla cómo se realizaron cada uno de estos pasos y las contribuciones finales de la tesis.
Novel conjunto de datos longitudinal------ Esta tesis contribuye con T1DiabetesGranada, un conjunto de datos longitudinales abierto que no sólo proporciona niveles continuos de glucemia en sangre, sino también información demográfica y clínica de los pacientes. El conjunto de datos incluye 257 780 días de mediciones que abarcan cuatro años de 736 pacientes con diabetes tipo 1 de la provincia de Granada, en España. Este conjunto de datos avanza más allá del estado del arte como uno de los conjuntos de datos abiertos más amplios temporalmente y extenso en mediciones continuas de glucemia en sangre, impulsando así el desarrollo de nuevos modelos de inteligencia artificial para la caracterización y predicción del nivel de glucemia en sangre.
Comprensión de los métodos de evaluación de los modelos de predicción------ Se desarrollan modelos de predicción generales basados en redes neuronales artificiales para la predicción de glucemia, los cuales son entrenados y evaluados utilizando datos continuos de glucemia de todos los pacientes. Se incluyen 3 algoritmos de predicción, una regresión lineal, así como arquitecturas de redes neuronales más avanzadas como LSTM y CNN. La predicción se realiza a partir de dos horas de valores históricos tomados cada 15 minutos, con 6 diferentes horizontes de predicción de 30 a 240 minutos. Para evaluar su rendimiento se usan tanto métricas clínicas (análisis de la rejilla de error de Clarke) como no clínicas (RMSE, MSE, MAE y MAPE), aplicadas no solo en todo el rango de glucemia, sino también independientemente en rangos clínicamente relevantes. Los resultados muestran un desempeño ligeramente mejor de los modelos de predicción lineales frente a los más complejos (LSTM y CNN), probablemente debido a la concentración de datos en un rango reducido de glucemia (100-200 mg/dL), lo que favorece a los modelos de predicción más simples. Se concluye que un enfoque holístico es clave para evaluar correctamente estos modelos de predicción. Visualizaciones como la rejilla de error de Clarke permiten analizar la geometría de las predicciones, mientras que evaluar las métricas en rangos específicos de glucemia evita conclusiones sesgadas y refleja con mayor precisión el comportamiento de los modelos de predicción en rangos de hipoglucemia e hiperglucemia que son los de más riesgo para la salud.
Identificación de grupo de pacientes------ Un aspecto clave de la tesis es identificar grupos de pacientes con patrones similares de glucemia mediante algoritmos de aprendizaje no supervisados como K-means, Gaussian Mixture, agrupamiento jerárquico y mapas autoorganizados. Se utilizan características extraídas de las mediciones de glucemia (promedio, desviación estándar, mínimo y máximo) en diferentes momentos del día para representar a cada paciente en el proceso de agrupamiento. Para estimar el número de grupos, se aplican el método del codo y el análisis de dendrogramas. Los resultados más prometedores se seleccionan, utilizando el coeficiente de Silhouette y un método personalizado, para presentarlos a expertos clínicos. Finalmente, se identifican seis grupos con patrones diferenciados de glucemia, que no solo sirven para desarrollar modelos de predicción basados en ellos, sino que también pueden influir en la atención médica. Estos grupos brindan la oportunidad a los endocrinos de diseñar tratamientos personalizados según los perfiles de pacientes. Con el uso de monitores continuos de glucemia, es posible asignar relativamente rápido a nuevos pacientes a uno de estos grupos, facilitando un enfoque terapéutico más dirigido desde las primeras fases de monitorización. Además, el proceso es aplicable a pacientes con historiales más largos de monitoreo, siempre que sus datos sean accesibles Evaluación de los modelos de predicción basados en grupos------ Una vez definidos los grupos de pacientes, se procede a crear modelos de predicción de los niveles de glucemia en sangre para cada grupo utilizando el mismo procedimiento de experimentación y modelos de predicción utilizado para los modelos generales. Para tener un punto de comparación adicional, se crean modelos de predicción personalizados para una muestra de pacientes, se seleccionan a aquellos cuyos niveles de glucemia fueron mejor y peor predichos por los modelos de predicción basados en grupos. Estos modelos de predicción personalizados entrenaron individualmente con los datos de cada paciente, replicando el procedimiento de entrenamiento aplicado a los modelos de predicción generales y basados en grupos. Los resultados muestran que los modelos de predicción basados en grupos superan a los modelos de predicción generales en ciertos grupos de pacientes y condiciones específicas, como en los grupos que representan a pacientes con un mejor control de glucemia. En estos casos las mejoras en métricas clínicas que evalúan el porcentaje de mediciones dentro de áreas clínicamente aceptable son más pronunciadas a medida que aumenta el horizonte de predicción, lo que sugiere que estos modelos pueden capturar patrones a largo plazo de manera más efectiva. Si bien los avances en los rangos de hipoglucemia son menores, se observaron mejoras en algunos grupos de pacientes con más eventos hipoglucémicos sobre todo en los algoritmos de predicción LSTM y CNN. Al comparar los modelos de predicción personalizados con los basados en grupos, aunque los primeros ofrecen mejoras en ciertos casos, no siempre superan a los modelos de predicción basados en grupos, mejorando solo en el 37.98% de las ocasiones. Por lo que se puede decir que los modelos de predicción basados en grupos ofrecen un equilibrio entre precisión y practicidad, mejorando el rendimiento sobre muchos de los modelos de predicción generales y requiriendo menos datos para la utilización en nuevos pacientes, siendo una solución eficaz para la predicción de glucemia en pacientes con diabetes tipo 1.
Conclusiones------ Los modelos de predicción basados en grupos resultan ser una solución intermedia prometedora entre los modelos generales y personalizados para la predicción de glucemia en pacientes con diabetes tipo 1. Estos modelos ofrecen mejoras sobre los modelos de predicción generales, particularmente en pacientes con mejor control de la glucemia y proporción elevada de hiperglucemias, equilibrando precisión y practicidad, y requiriendo menos datos para su implementación en nuevos pacientes. Aunque los modelos de predicción personalizados mostraron mejores resultados en algunos casos, los basados en grupos demostraron ser más eficaces en la mayoría de situaciones, consolidándose como una alternativa viable para optimizar la predicción y atención médica personalizada en el manejo de la diabetes tipo 1. Las principales contribuciones de la tesis son: 1.Un extenso conjunto de datos longitudinal multimodal de mediciones de glucemia continua, recopiladas utilizando medidores de glucemia continua, con información demográfica y clínica de los pacientes con diabetes tipo 1.
2.Una metodología de experimentación utilizando algoritmos de predicción aplicados a un conjunto de datos longitudinal extenso de mediciones de glucemia continua.
3.Tres conjuntos de modelos de predicción de niveles de glucemia continua basados en redes neuronales artificiales: generales, basados en grupos y personalizados.
4.Una metodología para crear grupos de pacientes basada en sus mediciones de glucemia continua.
5.Varios grupos de pacientes con diabetes tipo 1 que permiten a los endocrinos diseñar tratamientos personalizados según los perfiles únicos de los pacientes.
6.Una base de código para apoyar todos los procesos experimentales de las metodologías y la creación del conjunto de datos, permitiendo la reproducibilidad y la escalabilidad para futuras investigaciones.
Type 1 diabetes is primarily characterized by a severe issue in insulin secretion. Patients with this condition must rely on external sources of insulin to regulate their blood glucose levels, a task complicated by the high variability in glucose patterns across different individuals. These variations are influenced by lifestyle, metabolism, and sex, among others. Consequently, it is common for patients with type 1 diabetes to use continuous glucose monitoring devices to better manage their condition, as they provide real-time information about blood glucose levels. However, having an application that predicts future glucose levels would be valuable, helping patients take preventative measures to avoid complications arising from out-of-range glucose levels. Predicting these levels is inherently complex. On the one hand, general prediction models, which use continuous glucose data from multiple patients for training and evaluation, fail to fully capture the individual differences in glucose patterns. On the other hand, developing personalized prediction models for each patient is impractical, as it requires a large amount of continuous glucose data, meaning new patients would have to wait for an extended period before receiving accurate predictions. Additionally, there are very few open-access datasets with a sufficient number of patients and long-term continuous glucose measurements to support robust experimentation.
In this context, the goal of this thesis is to investigate intelligent techniques to support the management of type 1 diabetes mellitus in real-world settings by leveraging continuous glucose measurement data collected from mobile and wearable devices. To achieve this main goal, the following specific objectives are set: create an extensive, long-term, open-access longitudinal dataset consisting of blood glucose level measurements collected through continuous glucose monitoring sensors to develop new intelligent models; comprehend the capacity of existing evaluation methods for estimating the performance of blood glucose level prediction models in real-world settings, identify patient groups with common characteristics based on patients' longitudinal blood glucose level measurements; and evaluate the performance of group-based blood glucose level prediction models in real-world settings. For this evaluation, three types of prediction models will be developed and compared: general, group-based, and personalized prediction models, using both clinical and non-clinical evaluation metrics. This will determine the feasibility of the proposed approach for its use in blood glucose prediction in realistic settings.
To address this problem, a dataset with a sufficient number of continuous glucose measurements and an adequate number of patients is first collected, ensuring a solid foundation for the proposed experimentation. Next, general blood glucose prediction models are developed, where training and evaluation are performed using continuous glucose data from all patients in the dataset. This allowed for the identification of the most appropriate evaluation strategies for these types of prediction models. Once these strategies are understood, patient groups are created using various unsupervised learning algorithms. Given the various configurations tested, group sizes, dataset versions, and algorithms, the clustering results are extensive. The most promising results are selected using an internal clustering evaluation metric, supplemented by a procedure developed in this thesis, which clinical experts then review to identify the most appropriate solution. Once the groups are defined, the same experimentation for general prediction models is replicated within the groups to evaluate the comparative performance between the general and group-based prediction models. Additionally, patients within each group are selected to develop personalized prediction models, allowing further comparison with the group-based prediction models and determining whether the latter represents a viable intermediate solution between general and personalized prediction models. Each of these steps and the final contributions of the thesis are detailed below.
A novel longitudinal dataset ------ This thesis contributes with T1DiabetesGranada, an open longitudinal dataset that provides continuous blood glucose measurements and includes demographic and clinical information about the patients. The dataset comprises 257 780 days of measurements collected over four years from 736 type 1 diabetes patients in Granada, Spain. This dataset surpasses the current state of the art by being one of the most temporally extensive and comprehensive open datasets of continuous blood glucose measurements, driving the development of new artificial intelligence models for characterizing and predicting blood glucose levels.
Understanding evaluation methods for prediction models ------ General prediction models based on artificial neural networks are developed for blood glucose prediction, and they are trained and tested using continuous glucose data from all patients. These included prediction algorithms like linear regression and more advanced neural network architectures such as LSTM and CNN. Predictions are made using two hours of historical values taken every 15 minutes, with six different prediction horizons ranging from 30 to 240 minutes. Clinical metrics (such as Clarke error grid analysis) and non-clinical metrics (RMSE, MSE, MAE, and MAPE) are used to evaluate the prediction models' performance, applied to the full glucose range and independently to clinically relevant ranges. The results showed slightly better performance from linear prediction algorithms than more complex prediction algorithms (LSTM and CNN), likely due to the concentration of data in a narrow glucose range (100-200 mg/dL), favoring simpler prediction models. The conclusion is that a holistic approach is key to properly evaluating these prediction models. Visualizations such as Clarke error grids help analyze the geometry of the predictions while evaluating metrics within specific glucose ranges, prevent biased conclusions, and more accurately reflect the prediction models' behavior in hypoglycemic and hyperglycemic ranges, which pose the most significant health risks.
Patient group identification ------ A key aspect of this thesis is identifying groups of patients with similar glucose patterns using unsupervised learning algorithms such as K-means, Gaussian Mixture Models, hierarchical clustering, and self-organizing maps. Features extracted from continuous glucose measurements (mean, standard deviation, minimum, and maximum) at different times of the day are used to represent each patient during the clustering process. The elbow method and dendrogram analysis are applied to estimate the number of groups. The most promising results are selected using the Silhouette Coefficient and a customized method and are then presented to clinical experts. Ultimately, six groups with distinct glucose patterns are identified, which not only aid in the development of prediction models but also have the potential to influence medical care. These groups provide endocrinologists with the opportunity to design personalized treatments based on patient profiles. With continuous glucose monitors, new patients can be relatively quickly assigned to one of these groups, facilitating a more targeted therapeutic approach from the early stages of monitoring. The process also applies to patients with longer monitoring histories, provided their data is accessible.
Group-based prediction model evaluation ------ Once the patient groups are defined, blood glucose prediction models are developed for each group using the same experimental procedure and prediction algorithms applied to the general prediction models. To provide an additional point of comparison, personalized prediction models are created for a sample of patients, selecting those whose blood glucose is best and worst predicted by the group-based prediction models. These personalized prediction models are individually trained with each patient's data, replicating the training procedure applied to the general and group-based prediction models. The results showed that group-based prediction models outperformed general prediction models for certain patient groups and under specific conditions, such as in groups representing patients with better glucose control. In these cases, improvements in clinical metrics assessing the percentage of measurements within clinically acceptable zones became more pronounced as the prediction horizon increased, suggesting that these prediction models can capture long-term patterns more effectively. While improvements in hypoglycemic ranges are less significant, some improvements are observed in groups of patients with more frequent hypoglycemic events, particularly with LSTM and CNN prediction models. When comparing personalized prediction models to group-based prediction models, although the former offered improvements in some instances, they did not always outperform the group-based prediction models, with improvements seen in only 37.98% of cases. Therefore, group-based prediction models offer a balance between accuracy and practicality, outperforming many general prediction models and requiring less data for use in new patients, making them an effective solution for glucose prediction in type 1 diabetes patients.
Conclusions ------ Group-based prediction models are a promising intermediate solution between general and personalized prediction models for predicting blood glucose levels in type 1 diabetes patients. These prediction models outperform general prediction models, particularly for patients with better glucose control and higher proportions of hyperglycemia, balancing accuracy and practicality while requiring less data for implementation in new patients. Although personalized prediction models performed better in certain cases, group-based prediction models proved more effective in most situations, establishing themselves as a viable alternative for optimizing glucose prediction and providing personalized medical care in managing type 1 diabetes. The key contributions of this thesis include: 1.A large, multimodal longitudinal dataset of continuous glucose measurements collected using continuous glucose monitors, along with demographic and clinical information about type 1 diabetes patients.
2.A robust experimental methodology for applying prediction models to an extensive longitudinal dataset of continuous glucose measurements.
3.Three sets of prediction models for continuous glucose levels using artificial neural networks, namely general, group-based, and personalized prediction models.
4.A methodology for creating patient groups based on their continuous glucose measurements.
5.Various patient groups that enable endocrinologists to design personalized treatments based on unique patient profiles.
6.A codebase to support all experimental methodologies processes and the creation of the dataset, enabling reproducibility and scalability for future research.
- English
Type 1 diabetes is primarily characterized by a severe issue in insulin secretion. Patients with this condition must rely on external sources of insulin to regulate their blood glucose levels, a task complicated by the high variability in glucose patterns across different individuals. These variations are influenced by lifestyle, metabolism, and sex, among others. Consequently, it is common for patients with type 1 diabetes to use continuous glucose monitoring devices to better manage their condition, as they provide real-time information about blood glucose levels. However, having an application that predicts future glucose levels would be valuable, helping patients take preventative measures to avoid complications arising from out-of-range glucose levels. Predicting these levels is inherently complex. On the one hand, general prediction models, which use continuous glucose data from multiple patients for training and evaluation, fail to fully capture the individual differences in glucose patterns. On the other hand, developing personalized prediction models for each patient is impractical, as it requires a large amount of continuous glucose data, meaning new patients would have to wait for an extended period before receiving accurate predictions. Additionally, there are very few open-access datasets with a sufficient number of patients and long-term continuous glucose measurements to support robust experimentation.
In this context, the goal of this thesis is to investigate intelligent techniques to support the management of type 1 diabetes mellitus in real-world settings by leveraging continuous glucose measurement data collected from mobile and wearable devices. To achieve this main goal, the following specific objectives are set: create an extensive, long-term, open-access longitudinal dataset consisting of blood glucose level measurements collected through continuous glucose monitoring sensors to develop new intelligent models; comprehend the capacity of existing evaluation methods for estimating the performance of blood glucose level prediction models in real-world settings, identify patient groups with common characteristics based on patients' longitudinal blood glucose level measurements; and evaluate the performance of group-based blood glucose level prediction models in real-world settings. For this evaluation, three types of prediction models will be developed and compared: general, group-based, and personalized prediction models, using both clinical and non-clinical evaluation metrics. This will determine the feasibility of the proposed approach for its use in blood glucose prediction in realistic settings.
To address this problem, a dataset with a sufficient number of continuous glucose measurements and an adequate number of patients is first collected, ensuring a solid foundation for the proposed experimentation. Next, general blood glucose prediction models are developed, where training and evaluation are performed using continuous glucose data from all patients in the dataset. This allowed for the identification of the most appropriate evaluation strategies for these types of prediction models. Once these strategies are understood, patient groups are created using various unsupervised learning algorithms. Given the various configurations tested, group sizes, dataset versions, and algorithms, the clustering results are extensive. The most promising results are selected using an internal clustering evaluation metric, supplemented by a procedure developed in this thesis, which clinical experts then review to identify the most appropriate solution. Once the groups are defined, the same experimentation for general prediction models is replicated within the groups to evaluate the comparative performance between the general and group-based prediction models. Additionally, patients within each group are selected to develop personalized prediction models, allowing further comparison with the group-based prediction models and determining whether the latter represents a viable intermediate solution between general and personalized prediction models. Each of these steps and the final contributions of the thesis are detailed below.
A novel longitudinal dataset ------ This thesis contributes with T1DiabetesGranada, an open longitudinal dataset that provides continuous blood glucose measurements and includes demographic and clinical information about the patients. The dataset comprises 257 780 days of measurements collected over four years from 736 type 1 diabetes patients in Granada, Spain. This dataset surpasses the current state of the art by being one of the most temporally extensive and comprehensive open datasets of continuous blood glucose measurements, driving the development of new artificial intelligence models for characterizing and predicting blood glucose levels.
Understanding evaluation methods for prediction models ------ General prediction models based on artificial neural networks are developed for blood glucose prediction, and they are trained and tested using continuous glucose data from all patients. These included prediction algorithms like linear regression and more advanced neural network architectures such as LSTM and CNN. Predictions are made using two hours of historical values taken every 15 minutes, with six different prediction horizons ranging from 30 to 240 minutes. Clinical metrics (such as Clarke error grid analysis) and non-clinical metrics (RMSE, MSE, MAE, and MAPE) are used to evaluate the prediction models' performance, applied to the full glucose range and independently to clinically relevant ranges. The results showed slightly better performance from linear prediction algorithms than more complex prediction algorithms (LSTM and CNN), likely due to the concentration of data in a narrow glucose range (100-200 mg/dL), favoring simpler prediction models. The conclusion is that a holistic approach is key to properly evaluating these prediction models. Visualizations such as Clarke error grids help analyze the geometry of the predictions while evaluating metrics within specific glucose ranges, prevent biased conclusions, and more accurately reflect the prediction models' behavior in hypoglycemic and hyperglycemic ranges, which pose the most significant health risks.
Patient group identification ------ A key aspect of this thesis is identifying groups of patients with similar glucose patterns using unsupervised learning algorithms such as K-means, Gaussian Mixture Models, hierarchical clustering, and self-organizing maps. Features extracted from continuous glucose measurements (mean, standard deviation, minimum, and maximum) at different times of the day are used to represent each patient during the clustering process. The elbow method and dendrogram analysis are applied to estimate the number of groups. The most promising results are selected using the Silhouette Coefficient and a customized method and are then presented to clinical experts. Ultimately, six groups with distinct glucose patterns are identified, which not only aid in the development of prediction models but also have the potential to influence medical care. These groups provide endocrinologists with the opportunity to design personalized treatments based on patient profiles. With continuous glucose monitors, new patients can be relatively quickly assigned to one of these groups, facilitating a more targeted therapeutic approach from the early stages of monitoring. The process also applies to patients with longer monitoring histories, provided their data is accessible.
Group-based prediction model evaluation ------ Once the patient groups are defined, blood glucose prediction models are developed for each group using the same experimental procedure and prediction algorithms applied to the general prediction models. To provide an additional point of comparison, personalized prediction models are created for a sample of patients, selecting those whose blood glucose is best and worst predicted by the group-based prediction models. These personalized prediction models are individually trained with each patient's data, replicating the training procedure applied to the general and group-based prediction models. The results showed that group-based prediction models outperformed general prediction models for certain patient groups and under specific conditions, such as in groups representing patients with better glucose control. In these cases, improvements in clinical metrics assessing the percentage of measurements within clinically acceptable zones became more pronounced as the prediction horizon increased, suggesting that these prediction models can capture long-term patterns more effectively. While improvements in hypoglycemic ranges are less significant, some improvements are observed in groups of patients with more frequent hypoglycemic events, particularly with LSTM and CNN prediction models. When comparing personalized prediction models to group-based prediction models, although the former offered improvements in some instances, they did not always outperform the group-based prediction models, with improvements seen in only 37.98% of cases. Therefore, group-based prediction models offer a balance between accuracy and practicality, outperforming many general prediction models and requiring less data for use in new patients, making them an effective solution for glucose prediction in type 1 diabetes patients.
Conclusions ------ Group-based prediction models are a promising intermediate solution between general and personalized prediction models for predicting blood glucose levels in type 1 diabetes patients. These prediction models outperform general prediction models, particularly for patients with better glucose control and higher proportions of hyperglycemia, balancing accuracy and practicality while requiring less data for implementation in new patients. Although personalized prediction models performed better in certain cases, group-based prediction models proved more effective in most situations, establishing themselves as a viable alternative for optimizing glucose prediction and providing personalized medical care in managing type 1 diabetes. The key contributions of this thesis include: 1.A large, multimodal longitudinal dataset of continuous glucose measurements collected using continuous glucose monitors, along with demographic and clinical information about type 1 diabetes patients.
2.A robust experimental methodology for applying prediction models to an extensive longitudinal dataset of continuous glucose measurements.
3.Three sets of prediction models for continuous glucose levels using artificial neural networks, namely general, group-based, and personalized prediction models.
4.A methodology for creating patient groups based on their continuous glucose measurements.
5.Various patient groups that enable endocrinologists to design personalized treatments based on unique patient profiles.
6.A codebase to support all experimental methodologies processes and the creation of the dataset, enabling reproducibility and scalability for future research.
- español
