Developing Novel Criteria to Classify ARDS Severity using a Machine Learning Approach

• Autores: Mohammed Gamal Sayed Abdelall
• Directores de la Tesis: David Riaño Ramos (dir. tes.)
• Lectura: En la Universitat Rovira i Virgili ( España ) en 2022
• Idioma: inglés
• Número de páginas: 96
• Tribunal Calificador de la Tesis: José Manuel Juarez Herrero (presid.) , Domènec Puig Valls (secret.) , José Ramón Alonso Viladot (voc.)
• Enlaces
  • Tesis en acceso abierto en: TDX
• Resumen
  • español

    Frente a las dificultades médicas para abordar adecuadamente los problemas de ARDS, tal como se reportan en múltiples publicaciones especializadas, en esta tesis planteamos la hipótesis de que el uso de tecnologías modernas de aprendizaje automático (ML) podría mejorar nuestro conocimiento y nuestra capacidad para predecir y abordar estos ARDS. cuestiones. Para lograr estos objetivos (i), propusimos una fórmula novedosa [PaO2/(FiO2xPEEP) or P/FPE] para PEEP≥5 y los valores de corte correspondientes para abordar la brecha de definición de Berlín para la gravedad del ARDS mediante el uso de enfoques ML. Examinamos los valores de P/FPE que delimitan los límites del ARDS leve, moderado y grave. Aplicamos ML para predecir la gravedad del ARDS después del inicio a lo largo del tiempo comparando los criterios actuales de PaO2/FiO2 de Berlín con P/FPE en tres escenarios diferentes, (ii) apuntamos a caracterizar el mejor escenario temprano durante los dos primeros días en la unidad de cuidados intensivos (ICU) para predecir la duración de la MV después del inicio del ARDS utilizando enfoques de ML, y (iii) validamos P/FPE como predictor de mortalidad en la ICU más allá del estado actual del arte utilizando umbrales de clasificación intuitivos basados en ML.

  • català

    Davant de les dificultats mèdiques per abordar correctament els problemes del ARDS, tal com es reporten en múltiples publicacions especialitzades, en aquesta tesi hem plantejat la hipòtesi que l'ús de tecnologies modernes d'aprenentatge automàtic (ML) podria millorar el nostre coneixement i la nostra capacitat per predir i abordar aquests ARDS. qüestions. Per assolir aquests objectius (i) vam proposar una fórmula nova [PaO2/(FiO2xPEEP) or P/FPE] per a PEEP≥5 i els valors de tall corresponents per abordar la bretxa de definició de Berlín per a la gravetat de l'ARDS mitjançant enfocaments ML. Es van examinar els valors de P/FPE que delimiten els límits de l'ARDS lleu, moderat i greu. Hem aplicat ML per predir la gravetat del ARDS després de l'aparició al llarg del temps comparant els criteris actuals de PaO2/FiO2 de Berlín amb P/FPE en tres escenaris diferents, (ii) vam tenir com a objectiu caracteritzar el millor escenari precoç durant els dos primers dies a la unitat de cures intensives (ICU) per predir la durada de la MV després de l'inici de l'ARDS mitjançant enfocaments de ML, i (iii) vam validar P/FPE com a predictor de la mortalitat de la ICU més enllà de l'estat actual de la tècnica mitjançant llindars de classificació intuïtius basats en ML.

  • English

    Acute respiratory distress syndrome (ARDS) is a noncardiogenic pulmonary edema, lung inflammation with hypoxemia, and decreased lung compliance. ARDS is a heterogeneous syndrome with a fatal outcome, a constellation of clinical and physiologic observations thought to represent a common pathology. Pathogenesis of ARDS remains elusive, and there is no gold standard diagnostic test. There is a lot of heterogeneity in ARDS diagnosis, the possibility that ARDS is, in fact, a collection of different diseases that have not yet been separately identified. In addition, the disease trajectory of patients within each ARDS category can impact outcome. Most ARDS patients require mechanical ventilation (MV).

    In front of the medical difficulties to properly address ARDS issues, as they are reported in multiple specialized publications, in this thesis we hypothesized that the use of modern machine learning (ML) technologies could improve our knowledge and our capacity to predict and address these ARDS issues. In order to achieve these objectives (i) we proposed a novel formula [PaO2/(FiO2xPEEP) or P/FPE] for PEEP≥5 and corresponding cut-off values to address Berlin’s definition gap for ARDS severity by using ML approaches. We examined P/FPE values delimiting the boundaries of mild, moderate, and severe ARDS. We applied ML to predict ARDS severity after onset over time by comparing current Berlin PaO2/FiO2 criteria with P/FPE under three different scenarios, (ii) we aimed at characterizing the best early scenario during the first two days in the intensive care unit (ICU) to predict MV duration after ARDS onset using ML approaches, and (iii) we validated P/FPE as a predictor of ICU mortality beyond the current state of the art using intuitive classification thresholds based on ML.

    We extracted clinical data from the first 3 ICU days after ARDS onset from the single-center MIMIC-III critical care database (MetaVision, 2008-2012) and the multicenter eICU Collaborative Research Database across the United States between 2014 and 2015. Disease progression in each database was tracked along the first 3 ICU days to assess ARDS severity. We included variables of arterial oxygenation and ventilator settings that were readily available in routine clinical practice to guarantee clinical relevance within a wide range of ARDS severity. Three robust ML techniques were implemented using Python 3.7: LightGBM, RF, and XGBoost.

    We proved that our novel P/FPE index to assess ARDS severity after onset over time is markedly better than current PaO2/FiO2 ratio. The best early MV duration prediction model was obtained with data captured in the 2nd day. ML models might have important implications for optimizing ICU resource utilization and high acute cost reduction of MV. Moreover, P/FPE index is a more sensitive predictor of ICU mortality over time than the PaO2/FiO2 ratio in all ARDS categories.

    ARDS is a predominantly clinical diagnosis, but there have been difficulties in agreeing on a standardized, universal definition. The PaO2/FiO2 ratio classifies the severity of ARDS based on the degree of the oxygenation deficit and within each of those categories, patients are assumed to be less heterogeneous. However, recent observations showed that mild ARDS is underappreciated because it had a high mortality rate. Since the current Berlin definition for ARDS does not account for PEEP in its calculation, it provides an incomplete picture of actual ARDS severity. This thesis provides a solution for that dilemma by using the P/FPE index, taking applied PEEP into account and creating three grades of severity based on intuitive classification thresholds that are different from the Berlin definition, and within each of these ARDS categories, the patients had a similar degree of lung severity. We believe this thesis is an important addition to the current ARDS story.