Prediction of heart failure decompensations using artificial intelligence - machine learning techniques

Resumen

Heart failure (HF) is a major concern in public health. Its total impact is increased by its high incidence and prevalence and its unfavourable medium-term prognosis. In addition, HF leads to huge health care resource consumption. Moreover, efforts to develop a deterministic understanding of rehospitalization have been difficult, as no specific patient or hospital factors have been shown to consistently predict 30-day readmission after hospitalization for HF.Taking all these facts into account, we wanted to develop a project to improve the assistance care of patients with HF. Up to know, we were using telemonitoring with a codification system that generated alarms depending on the received values. However, these simple rules generated large number of false alerts being, hence, not trustworthy. The final aims of this work are: (i) asses the benefits of remote patient telemonitoring (RPT), (ii) improve the results obtained with RPT using ML techniques, detecting which parameters measured by telemonitoring best predict HF decompensations and creating predictive models that will reduce false alerts and detect early decompensations that otherwise will lead to hospital admissions and (iii) determine the influence of environmental factors on HF decompensations.All in all, the conclusions of this study are:1. Asses the benefits of RPT: Telemonitoring has not shown a statistically significant reduction in the number of HF-related hospital admissions. Nevertheless, we have observed a statistically significant reduction in mortality in the intervention group with a considerable percentage of deaths from non-cardiovascular causes. Moreover, patients have considered the RPT programme as a tool that can help them in the control of their chronic disease and in the relationship with health professionals.2. Improve the results obtained with RPT using machine learning techniques: Significant weight increases, desaturation below 90%, perception of clinical worsening, including development of oedema, worsening of functional class and orthopnoea are good predictors of heart failure decompensation. In addition, machine learning techniques have improved the current alerts system implemented in our hospital. The system reduces the number of false alerts notably although it entails a decrement on sensitivity values. The best results are achieved with the predictive model built by applying NB with Bernoulli to the combination of telemonitoring alerts and questionnaire alerts (Weight + Ankle + well-being plus the yellow alerts of systolic blood pressure, diastolic blood pressure, O2Sat and heart rate). 3. Determine the influence of environmental factors on HF decompensations: Air temperature is the most significant environmental factor (negative correlation) in our study, although some other attributes, such as precipitation, are also relevant. This work also shows a consistent association between increasing levels SO2 and NOX air and HF hospitalizations.