A machine-learning-based model for the one-year mortality prediction in patients admitted to an intensive care unit with diagnosis of sepsis

• Autores: Javier Esteban García Gallo
• Directores de la Tesis: Jhon Freddy Duitama Muñoz (dir. tes.), Nelson Javier Fonseca Ruiz (dir. tes.)
• Lectura: En la Universidad de Antioquia ( Colombia ) en 2019
• Idioma: inglés
• Enlaces
  • Tesis en acceso abierto en: hdl.handle.net
• Resumen

  • The Intensive Care Unit (ICU) is a hospital department that provides intensive treatment to patients with severe and life-threatening conditions. The primary function of the ICU is to deliver care which cannot be administered in other areas of the hospital. Patients in the ICU are the most heavily monitored patients in the entire hospital; for this reasons the ICU is a data rich environment, even to the point of exhaustion. The vast amount of data obtained from a single patient in an intensive care unit makes it humanly impossible to organize and interpret it in the required time, thus, scores that model the patient severity and can be related with the mortality have been created. The primary motivation of this scores was to derive further insight of the patient condition and improve patient care. Traditionally, this scores are population-based and provide statistically rigorous results for an average patient, and are useful to guide prognostication, to assess ongoing disease development and organ function, to compare ICU performance over time and across units and to compare clinical trial population outcomes but, pitifully, they are not precise enough to draw conclusions about groups of patients that share a relevant clinical condition, like a particular disease, and even less to be used for individual prediction of outcomes. When standard scores do not fit the data of a specific population well enough, two approaches to adapting them for use among patients with the specific condition have been used. One approach would be to modify the traditional score by adapting them for use specifically among patients that share a condition, which we will be referring as adjusted models. The other approach would be to develop entirely new models based on a population that shares a common characteristics and that incorporates additional variables that could potentially enhance accuracy, which we will be referring as customized models. Sepsis patients are a specific population that is especially vulnerable, since they present a high in-hospital mortality of 25–30% and patients with sepsis are frequently cared for in ICUs, either because sepsis itself led to their admission or because sepsis developed as a complication of their admission for other reasons; moreover, it has been reported that sepsis survivors had substantially increased risks of all-cause mortality, as well as major health complications at 1 year after discharge when compared with the general population. For sepsis patients within the ICU, mortality prediction has been accessed through both adjusted and customized models; however, approaches addressed so far have focused on the in-hospital mortality prediction, and no methods have been proposed to identify and predict long-term risk and mortality in sepsis patients that are being taken care of in the ICU. According to the above, in this work, we present the development of a model that goes beyond the prediction of in-hospital mortality and alert those patients who may have a poor prognosis after being discharged from the hospital, and we formulate our research question as follows: Among adult ICU patients, is it possible to identify those who are at risk of dying one year after their sepsis related admission using demographic variables, comorbidities and physiological data obtained during the first 24 hours of their ICU stay? In order to answer such a question, we used three approaches. First we developed a custom one-year mortality prediction model using a Stochastic Gradient Boosting (SGB) technique. The model was based on the data of 5650 ICU patient’s admissions that were retrospectively identified as having sepsis, and used 132 predictors, obtained from variables found in the literature review or suggested by experts. In the first approach, we also used two techniques to measure the importance of the used predictors, and we found 17 predictors that allowed us to develop an SGB model with a performance similar to the complete model (which uses all the 132 predictors).

    In the second approach, we developed a methodology that allows the stratification of patients according to their one-year mortality risk. For this, we extended our study cohort using two additional retrospective criteria for sepsis identification and focusing only on the variables that were relevant (according to the results of the previous approach) or that were routinely taken to patients within the ICU, obtaining 15082 admissions; From said cohort we developed two scores systems that are correlated with the one-year mortality risk of the patients. Although the developed customized models for sepsis patient within the ICU proved far outperform adjusted scores for the one-year mortality prediction task, they continue to be population-based and therefore they provide “the average best choice” for sepsis patients. For this reason, in the third approach, we also propose and evaluate the generation of personalized models based on patient similarity metrics. The goal of this personalized models is to identify patients who are similar to a new patient and derive insights from the data of those similar patients to provide personalized predictions.

    Personalized models has been widely used for predictions in several fields, including music, movies and e-commerce, however, there are still very few studies that focus on personalized prediction models based on health data prediction. Moreover, no studies have been reported in which personalized models are developed from a population known to be very homogenous, such as our study population, where it is known that all patients have infection, organ dysfunction, and ICU stays of more than 24 hours. The developed models, with the three approaches, showed discrimination superior to adjusted models based on traditional severity scores and, the population based methodologies also presented adequate calibration. Specifically, our personalized models demonstrated the value of patient similarity metrics in outcome prediction modeling and showed superiority when compared to population-based models. Also, since we focused on long-term mortality prediction, these models successfully identify those patients who are at risk of dying one year after their sepsis related admission using demographic variables, comorbidities and physiological data obtained during the first 24 hours of their ICU stay, indicating early, which patients should be accompanied, observed attentively and provided with additional care that improve their quality of life. Finally, in order to enable the clinical use of the machine learning models developed for the prediction of one-year mortality of sepsis patients within the ICU, we developed a software based on the models that presented a better performance and the functionalities that are considered useful so that intensivist can obtain details of the particular condition of each patient and provide better care.