Search

En este trabajo se presenta una contribución en el diseño y desarrollo de soluciones de Inteligencia Artificial Confiables basadas en modelos predictivos en el ámbito de la salud. En el año 2019, la Unión Europea comenzó a marcar, a través de sus directrices, el camino que el diseño y desarrollo de soluciones basadas en IA deben seguir con el objetivo de promover la adopción de esta tecnología a una mayor escala y con las suficientes garantías de respeto de los derechos humanos, la ética y la seguridad en su aplicación. Estas directrices buscan impulsar un uso responsable y confiable de la IA en diversos ámbitos, entre los cuales se destaca el ámbito de la salud. En este contexto, en esta tesis se han explorado metodologías, técnicas y estrategias que promuevan el diseño y desarrollo de sistemas de IA confiables. Uno de los pilares fundamentales de esta tesis ha sido el uso de la metodología de diseño centrado en la persona con el propósito de que las soluciones de IA propuestas estén alineadas con las necesidades y expectativas y contextos de los usuarios finales. Por otro lado, el uso de modelos altamente interpretables y la propuesta de una innovadora estrategia de validación demuestra el compromiso con la investigación en el diseño y desarrollo sistemas de IA confiables en el ámbito de la salud. La tesis se ha desarrollado en el marco de casos de uso reales, en los que se han contado con el conocimiento experto de un equipo de profesionales clínicos, cuya participación ha permitido proponer soluciones que puedan contribuir positivamente al campo de la salud y la asistencia clínica., This work presents a contribution to the design and development of Trustworthy Artificial Intelligence solutions based on predictive models in the healthcare domain. In 2019, the European Union started to establish a path for the design and development of AI-based solutions through its guidelines. The aim is to promote a more widespread adoption of this technology, ensuring it adheres to human rights, ethics and security in its application. These guidelines seek to drive a responsible and trustworthy use of AI in healthcare, among other domains. In this context, different methodologies, techniques and strategies that promote the design and development of trustworthy AI have been explored in this work. On one hand, one of the main foundations of this thesis has been the use of a usercentered design methodology with the objective to ensure that the proposed AI solutions are aligned with the needs, expectations, and contexts of the end users. On the other hand, the use of highly interpretable models and the innovative validation strategy proposed demonstrate the commitment in designing and developing trustworthy AI systems in healthcare. This thesis has been conducted within the context of real-world use cases with the input of a group of clinical experts. Their participations has enabled the proposal of solutions that could make a positive impact in the healthcare domain and clinical support.

Sharing healthcare data is increasingly essential for developing data-driven improvements in patient care at the Intensive Care Unit (ICU). However, it is also very challenging under the strict privacy legislation of the European Union (EU). Therefore, we explored four successful open ICU healthcare databases to determine how open healthcare data can be shared appropriately in the EU. A questionnaire was constructed based on the Delphi method. Then, follow-up questions were discussed with experts from the four databases. These experts encountered similar challenges and regarded ethical and legal aspects to be the most challenging. Based on the approaches of the databases, expert opinion, and literature research, we outline four distinct approaches to openly sharing healthcare data, each with varying implications regarding data security, ease of use, sustainability, and implementability. Ultimately, we formulate seven recommendations for sharing open healthcare data to guide future initiatives in sharing open healthcare data to improve patient care and advance healthcare.

Purpose Sepsis is a leading cause of morbidity and mortality worldwide and is characterized by vascular leak. Treatment for sepsis, specifically intravenous fluids, may worsen deterioration in the context of vascular leak. We therefore sought to quantify vascular leak in sepsis patients to guide fluid resuscitation. Methods We performed a retrospective cohort study of sepsis patients in four ICU databases in North America, Europe, and Asia. We developed an intuitive vascular leak index (VLI) and explored the relationship between VLI and in-hospital death and fluid balance using generalized additive models (GAM). Results Using a GAM, we found that increased VLI is associated with an increased risk of in-hospital death. Patients with a VLI in the highest quartile (Q4), across the four datasets, had a 1.61–2.31 times increased odds of dying in the hospital compared to patients with a VLI in the lowest quartile (Q1). VLI Q2 and Q3 were also associated with increased odds of dying. The relationship between VLI, treated as a continuous variable, and in-hospital death and fluid balance was statistically significant in the three datasets with large sample sizes. Specifically, we observed that as VLI increased, there was increase in the risk for in-hospital death and 36–84 h fluid balance. Conclusions Our VLI identifies groups of patients who may be at higher risk for in-hospital death or for fluid accumulation. This relationship persisted in models developed to control for severity of illness and chronic comorbidities.

The unprecedented global crisis brought about by the COVID-19 pandemic has sparked numerous efforts to create predictive models for the detection and prognostication of SARS-CoV-2 infections with the goal of helping health systems allocate resources. Machine learning models, in particular, hold promise for their ability to leverage patient clinical information and medical images for prediction. However, most of the published COVID-19 prediction models thus far have little clinical utility due to methodological flaws and lack of appropriate validation. In this paper, we describe our methodology to develop and validate multi-modal models for COVID-19 mortality prediction using multi-center patient data. The models for COVID-19 mortality prediction were developed using retrospective data from Madrid, Spain (N = 2547) and were externally validated in patient cohorts from a community hospital in New Jersey, USA (N = 242) and an academic center in Seoul, Republic of Korea (N = 336). The models we developed performed differently across various clinical settings, underscoring the need for a guided strategy when employing machine learning for clinical decision-making. We demonstrated that using features from both the structured electronic health records and chest X-ray imaging data resulted in better 30-day mortality prediction performance across all three datasets (areas under the receiver operating characteristic curves: 0.85 (95% confidence interval: 0.83–0.87), 0.76 (0.70–0.82), and 0.95 (0.92–0.98)). We discuss the rationale for the decisions made at every step in developing the models and have made our code available to the research community. We employed the best machine learning practices for clinical model development. Our goal is to create a toolkit that would assist investigators and organizations in building multi-modal models for prediction, classification, and/or optimization.

High resolution clinical databases from electronic health records are increasingly being used in the field of health data science. Compared to traditional administrative databases and disease registries, these newer highly granular clinical datasets offer several advantages, including availability of detailed clinical information for machine learning and the ability to adjust for potential confounders in statistical models. The purpose of this study is to compare the analysis of the same clinical research question using an administrative database and an electronic health record database. The Nationwide Inpatient Sample (NIS) was used for the low-resolution model, and the eICU Collaborative Research Database (eICU) was used for the high-resolution model. A parallel cohort of patients admitted to the intensive care unit (ICU) with sepsis and requiring mechanical ventilation was extracted from each database. The primary outcome was mortality and the exposure of interest was the use of dialysis. In the low resolution model, after controlling for the covariates that are available, dialysis use was associated with an increased mortality (eICU: OR 2.07, 95% CI 1.75–2.44, p<0.01; NIS: OR 1.40, 95% CI 1.36–1.45, p<0.01). In the high-resolution model, after the addition of the clinical covariates, the harmful effect of dialysis on mortality was no longer significant (OR 1.04, 95% 0.85–1.28, p = 0.64). The results of this experiment show that the addition of high resolution clinical variables to statistical models significantly improves the ability to control for important confounders that are not available in administrative datasets. This suggests that the results from prior studies using low resolution data may be inaccurate and may need to be repeated using detailed clinical data.

Objectives: Derangements of chloride ion concentration ([Cl–]) have been shown to be associated with acute kidney injury and other adverse outcomes. For a physicochemical approach, however, chloride ion concentration should be considered with sodium ion concentration. This study aimed to examine the association of chloride ion concentration and the main strong ion difference (difference between sodium ion concentration and chloride ion concentration) during the first 24 hours after admission into ICU with the development of acute kidney injury and mortality. Design: Retrospective analyses using the eICU Collaborative Research Database. Setting: ICUs in 208 hospitals across the United States between 2014 and 2015. Patients: Critically ill patients who were admitted into the ICU. Interventions: None. Measurements and Main Results: A total of 34,801 patients records were analyzed. A multivariable logistic regression analysis for the development of acute kidney injury within 7 days of ICU admission shows that, compared with main strong iron difference 32–34 mEq/as a reference, there were significantly high odds for the development of acute kidney injury in nearly all groups with main strong iron difference more than 34 mEq/L (main strong iron difference = 34–36 mEq/L, odds ratio = 1.17, p = 0.02; main strong iron difference = 38–40 mEq/L, odds ratio = 1.40, p < 0.001; main strong iron difference = 40–42 mEq/L, odds ratio = 1.46, p = 0.001; main strong iron difference > 42 mEq/L, odds ratio = 1.56, p < 0.001). With chloride ion concentration 104–106 mEq/L as a reference, the odds for acute kidney injury were significantly higher only in chloride ion concentration less than or equal to 94 mEq/L and chloride ion concentration 98–100 mEq/L groups. Analyses conducted using inverse probability weighting showed significantly greater odds for ICU mortality in all groups with main strong iron difference greater than 34mEq/L other than the 36–38mEq/L group, as well as in the less, National Institute of Health (Grant NIBIB R01 EB017205)

Illness severity scores are regularly employed for quality improvement and benchmarking in the intensive care unit, but poor generalization performance, particularly with respect to probability calibration, has limited their use for decision support. These models tend to perform worse in patients at a high risk for mortality. We hypothesized that a sequential modeling approach wherein an initial regression model assigns risk and all patients deemed high risk then have their risk quantified by a second, high-risk-specific, regression model would result in a model with superior calibration across the risk spectrum. We compared this approach to a logistic regression model and a sophisticated machine learning approach, the gradient boosting machine. The sequential approach did not have an effect on the receiver operating characteristic curve or the precision-recall curve but resulted in improved reliability curves. The gradient boosting machine achieved a small improvement in discrimination performance and was similarly calibrated to the sequential models., National Institutes of Health (U.S.) (Grant EB017205)

Objetivo. Evaluar los hábitos de estilo de vida y estado nutricional relacionados con las diferentes carreras universitarias. Materiales y métodos. Estudio descriptivo de corte transversal en 800 participantes; basados en una encuesta guía validada para el registro de datos a estudiantes de diferentes facultades de la Universidad Simón Bolívar de Barranquilla, Colombia. Se incluyeron datos básicos del estudiante, información sobre el estilo de vida y medidas antropométricas para así categorizarlos por IMC según la OMS. Se aplicaron criterios de inclusión y exclusión. Los datos fueron analizados estadísticamente (p<0.001). Resultados. Se excluyen 38, quedando un total de 762 estudiantes analizados y distribuidos de la siguiente manera; n=432 estudiantes en las Socio-sanitarias, n=131 estudiantes en Derecho y n=199 estudiantes en las Técnicas. El sexo femenino representa la mayoría de los participantes con un total de 430(56.3%) vs. 332(43.6%) de sexo masculino. En la facultad Sociosanitaria las mujeres representaron de igual manera la mayoría con un total de 269(62.3%) vs. 163(37.7%) de sexo masculino; así mismo ocurrió en la carrera universitaria de Derecho 86(65.6%) mujeres vs. 45(34.4%) de sexo masculino, caso contrario en las Técnicas, donde el sexo femenino represento a la población menor con 73(36.7%) vs. 126(63.3%) del sexo masculino; p<0.001. La edad promedio del total de estudiantes (n=762) fue de 19.2 +/- 2.3 años. En las Socio-sanitarias fue de 19.6+/-2.2 años, en Derecho fue de 19.4+/-3.1años y las Técnicas tuvieron el grupo menor de edad en años con 18.3+/-1.6 años; p<0.001. Con respecto al IMC, el promedio total de los estudiantes fue de 23.1+/.4.3 kg/m2, y su distribución en las diferentes carrera no fue significativo; p=0.866. En cuanto a la relación cintura/talla el promedio total de los estudiantes fue de 45.5+/-7.3cm, siendo la facultad de las Técnicas la que expreso un mayor valor con un promedio 47.1+/-5.7cm; p<0.001. La variable cadera/cintura t, Objective. To evaluate lifestyle habits and nutritional status related to different university careers. Material and methods. A descriptive cross-sectional study of 800 participants; based on a validated survey for the registration of the student´s data of different faculties of the Simón Bolívar University of Barranquilla, Colombia. Basic student data, lifestyle information and anthropometric measurements were included to categorize them by BMI according to the WHO. Inclusion and exclusion criteria were applied. The data were analyzed statistically (p <0.001). Results. 38 are excluded, leaving a total of 762 students analyzed and distributed as follows; n = 432 Health students, n = 131 Law students and n = 199 Techniques students. The female sex represents the majority of the participants with a total of 430 (56.3%) vs. 332 (43.6%) male. In the Health-careers faculty, women represented the same majority with a total of 269 (62.3%) vs. 163 (37.7%) male; likewise, it happened in the university career of Law 86 (65.6%) women vs. 45 (34.4%) of the masculine sex, otherwise in the Techniques, where the feminine sex represented the smaller population with 73 (36.7%) vs. 126 (63.3%) of the male sex; p<0.001. Totals average age was 19.2 +/- 2.3 years; in the health group was 19.6 +/- 2.2 years, in Law were 19.4 +/- 3.1 years and 18.3 +/- 1.6 years in Techniques; p<0.001. Regarding BMI, the total average was 23.1 +/- 4.3 kg /m2, and their distribution in the different races was not significant; p=0.866. As for the waist / height ratio, the total average of the students was 45.5 +/- 7.3cm, with the Technical faculty expressing a higher value with an average of 47.1 +/- 5.7cm; p<0.001. The hip / waist variable had a total average of 1.25 +/- 0.28, being the law career the one that expressed the lowest value with an average of 1.18 +/- 0.16 against the total population; p<0.001. Health students have a lower sleep habit, they sleep 2.9 +/- 0.7 hours per day; p<0.001. Conclusion.