Clinical decision support systems for palliative care referral: design and evaluation of frailty and mortality predictive models
- Autores: Vicent Blanes Selva
- Directores de la Tesis: Juan Miguel García Gómez (dir. tes.), Ascensión Doñate Martínez (dir. tes.)
- Lectura: En la Universitat Politècnica de València ( España ) en 2022
- Idioma: español
- Tribunal Calificador de la Tesis: Oscar Arturo Mayora Ibarra (presid.), Ángeles Calduch-Losa (secret.), José Francisco Romero Gómez (voc.)
- Programa de doctorado: Programa de Doctorado en Tecnologías para la Salud y el Bienestar por la Universitat Politècnica de València
- Materias:
- Enlaces
- Tesis en acceso abierto en: RiuNet
- Resumen
Palliative Care (PC) is specialized medical care that aims to improve patients' quality of life with serious illnesses. Historically, it has been applied to terminally ill patients, especially those with oncologic diagnoses. However, current research results suggest that PC positively affects the quality of life of patients with different conditions. The current trend on PC is to include non-oncological patients with conditions such as Chronic Obstructive Pulmonary Disease (COPD), organ function failure or dementia. However, the identification of patients with those needs is complex, and therefore alternative tools based on clinical data are required.
The growing demand for PC may benefit from a screening tool to identify patients with PC needs during hospital admission. Several tools, such as the Surprise Question (SQ) or the creation of different indexes and scores, have been proposed with varying degrees of success. Recently, the use of artificial intelligence algorithms, specifically Machine Learning (ML), has arisen as a potential solution given their capacity to learn from the Electronic Health Records (EHRs) and with the expectation to provide accurate predictions for admission to PC programs.
This thesis focuses on creating ML-based digital tools for identifying patients with palliative care needs at hospital admission. We have used mortality and frailty as the two clinical criteria for decision-making, being short survival and increased frailty, as our targets to make predictions. We also have focused on implementing these tools in clinical settings and studying their usability and acceptance in clinical workflows.
To accomplish these objectives, first, we studied and compared ML algorithms for one-year survival in adult patients during hospital admission. To do so, we defined a binary variable to predict, equivalent to the SQ and defined the set of predictive variables based on literature. We compared models based on Support Vector Machine (SVM), k-Nearest Neighbours (kNN), Random Forest (RF), Gradient Boosting Machine (GBM) and Multilayer Perceptron (MLP), attending to their performance, especially to the Area under the ROC curve (AUC ROC). Additionally, we obtained information on the importance of variables for tree-based models using the GINI criterion.
Second, we studied frailty measurement of Quality of Life (QoL) in candidates for PC intervention. For this second study, we narrowed the age of the population to elderly patients (¿ 65 years) as the target group. Then we created three different models: 1) for the adaptation of the one-year mortality model for elderly patients, 2) a regression model to estimate the number of days from admission to death to complement the results of the first model, and finally, 3) a predictive model for frailty status at one year. These models were shared with the academic community through a web application a that allows data input and shows the prediction from the three models and some graphs with the importance of the variables.
Third, we proposed a version of the 1-year mortality model in the form of an online calculator. This version was designed to maximize access from professionals by minimizing data requirements and making the software responsive to the current technological platforms. So we eliminated the administrative variables specific to the dataset source and worked on a process to minimize the required input variables while maintaining high the model's AUC ROC. As a result, this model retained most of the predictive power and required only seven bed-side inputs.
Finally, we evaluated the Clinical Decision Support System (CDSS) web tool on PC with an actual set of users. This evaluation comprised three domains: evaluation of participant's predictions against the ML baseline, the usability of the graphical interface, and user experience measurement. A first evaluation was performed, followed by a period of implementation of improvements and corrections to the pla
