Improving the compromise between accuracy, interpretability and personalization of rule-based machine learning in medical problems.

One of the key challenges when developing a predictive model is the capability to describe the domain knowledge and the cause-effect relationships in a simple way. Decision rules are a useful and important methodology in this context, justifying their application in several areas, particularly in clinical practice. Several machine-learning classifiers have exploited the advantageous properties of decision rules to build intelligent prediction models, namely decision trees and ensembles of trees (ETs). However, such methodologies usually suffer from a trade-off between interpretability and predictive performance. Some procedures consider a simplification of ETs, using heuristic approaches to select an optimal reduced set of decision rules. In this paper, we introduce a novel step to those methodologies. We create a new component to predict if a given rule will be correct or not for a particular patient, which introduces personalization into the procedure. Furthermore, the validation results using three public clinical datasets suggest that it also allows to increase the predictive performance of the selected set of rules, improving the mentioned trade-off.

A new approach for interpretability and reliability in clinical risk prediction: Acute coronary syndrome scenario.

INTRODUCTION: The risk prediction of the occurrence of a clinical event is often based on conventional statistical procedures, through the implementation of risk score models. Recently, approaches based on more complex machine learning (ML) methods have been developed. Despite the latter usually have a better predictive performance, they obtain little approval from the physicians, as they lack interpretability and, therefore, clinical confidence. One clinical issue where both types of models have received great attention is the mortality risk prediction after acute coronary syndromes (ACS). OBJECTIVE: We intend to create a new risk assessment methodology that combines the best characteristics of both risk score and ML models. More specifically, we aim to develop a method that, besides having a good performance, offers a personalized model and outcome for each patient, presents high interpretability, and incorporates an estimation of the prediction reliability which is not usually available. By combining these features in the same approach we expect that it can boost the confidence of physicians to use such a tool in their daily activity. METHODS: In order to achieve the mentioned goals, a three-step methodology was developed: several rules were created by dichotomizing risk factors; such rules were trained with a machine learning classifier to predict the acceptance degree of each rule (the probability that the rule is correct) for each patient; that information was combined and used to compute the risk of mortality and the reliability of such prediction. The methodology was applied to a dataset of 1111 patients admitted with any type of ACS (myocardial infarction and unstable angina) in two Portuguese hospitals, to assess the 30-days all-cause mortality risk, being validated through a Monte-Carlo cross-validation technique. The performance was compared with state-of-the-art approaches: logistic regression (LR), artificial neural network (ANN), and clinical risk score model (namely the Global Registry of Acute Coronary Events - GRACE). RESULTS: For the scenario being analyzed, the performance of the proposed approach and the comparison models was assessed through discrimination and calibration. The ability to rank the patients was evaluated through the area under the ROC curve (AUC), and the ability to stratify the patients into low or high-risk groups was determined using the geometric mean (GM) of specificity and sensitivity, the negative predictive value (NPV) and the positive predictive value (PPV). The validation calibration curves were also inspected. The proposed approach (AUC = 81%, GM = 74%, PPV = 17%, NPV = 99%) achieved testing results identical to the standard LR model (AUC = 83%, GM = 73%, PPV = 16%, NPV=99%), but offers superior interpretability and personalization; it also significantly outperforms the GRACE risk model (AUC = 79%, GM = 47%, PPV = 13%, NPV = 98%) and the standard ANN model (AUC = 78%, GM = 70%, PPV = 13%, NPV = 98%). The calibration curve also suggests a very good generalization ability of the obtained model as it approaches the ideal curve (slope = 0.96). Finally, the reliability estimation of individual predictions presented a great correlation with the misclassifications rate. CONCLUSION: We developed and described a new tool that showed great potential to guide the clinical staff in the risk assessment and decision-making process, and to obtain their wide acceptance due to its interpretability and reliability estimation properties. The methodology presented a good performance when applied to ACS events, but those properties may have a beneficial application in other clinical scenarios as well.

Multi-parametric prediction for cardiovascular risk assessment.

The employment of personal health systems (pHealth) is a valuable concept in the management of chronic diseases, particularly in the context of cardiovascular diseases. By means of a continuous monitoring of the patient it is possible to seamless access multiple sources of data, including physiological signals, providing professionals with a global and reliable view of the patient's status. In practice, it is possible the prompt diagnosis of events, the early prediction of critical events and the implementation of personalized therapies. Furthermore, the information collected during long periods creates new opportunities in the diagnosis of a disease, in its evolution, and in the prediction of possible complications. The focus of this work is the research and implementation of multi-parametric algorithms for data analysis in pHealth context, including data mining techniques as well as physiological signal modelling and processing. In particular, fusion strategies for cardiovascular status evaluation (namely cardiovascular risk assessment and cardiac function estimation) and multi-parametric prediction algorithms for the early detection of cardiovascular events (such as hypertension, syncope and heart failure decompensation) will be addressed.

New approaches for improving cardiovascular risk assessment.

INTRODUCTION AND OBJECTIVES: Clinical guidelines recommend the use of cardiovascular risk assessment tools (risk scores) to predict the risk of events such as cardiovascular death, since these scores can aid clinical decision-making and thereby reduce the social and economic costs of cardiovascular disease (CVD). However, despite their importance, risk scores present important weaknesses that can diminish their reliability in clinical contexts. This study presents a new framework, based on current risk assessment tools, that aims to minimize these limitations. METHODS: Appropriate application and combination of existing knowledge is the main focus of this work. Two different methodologies are applied: (i) a combination scheme that enables data to be extracted and processed from various sources of information, including current risk assessment tools and the contributions of the physician; and (ii) a personalization scheme based on the creation of patient groups with the purpose of identifying the most suitable risk assessment tool to assess the risk of a specific patient. RESULTS: Validation was performed based on a real patient dataset of 460 patients at Santa Cruz Hospital, Lisbon, Portugal, diagnosed with non-ST-segment elevation acute coronary syndrome. Promising results were obtained with both approaches, which achieved sensitivity, specificity and geometric mean of 78.79%, 73.07% and 75.87%, and 75.69%, 69.79% and 72.71%, respectively. CONCLUSIONS: The proposed approaches present better performances than current CVD risk scores; however, additional datasets are required to back up these findings.

Personalization Based on Grouping Strategies for Short-Term Cardiovascular Event Risk Assessment.

Cardiovascular diseases are the main cause of death in Europe, representing 47% of all deaths. This could be avoided, if each patient underwent the most adequate treatment. For this to happen, it is important to determine the patient's risk of having a cardiovascular event. This is known as risk assessment, and can be done using risk scores. However, there are several risk scores with similar performances, which makes it difficult to choose the most adequate one. We propose to overcome this by combining risk scores using personalization based on groups, where new patients are assigned to the most similar group and consequently to the most adequate risk score. This eliminates the need to choose a specific tool, and improves the overall performance (when compared with the performance of individual tools). This strategy was validated using the Santa Cruz Dataset. The results obtained were able to maintain the highest sensitivity while improving the specificity in 13% when compared with the highest values achieved by the selected individual risk scores (GRACE, TIMI, PURSUIT).

Correlation between autonomous function and left ventricular performance after acute myocardial infarction.

Reduced ejection fraction (EF), possibly induced/mediated by autonomic abnormal activation, is one of the most powerful predictors of adverse outcome after acute myocardial infarction (MI). A deep understanding of the correlation between the autonomous functionality and the left ventricular performance in these patients is therefore of paramount importance. The autonomous function is reflected in the cardiac activity and, specifically, in the heart rate variability (HRV) signal. Given the cardiac activity nonlinearity, growing interest is being manifested towards nonlinear methods of analysis, which might provide more significant information than the traditional linear approaches. The aim of the present study was to investigate if non-linear HRV metrics change between MI patients with preserved EF (pEF) and MI patients with reduced EF (rEF). Data were acquired in the context of the cardioRisk project. Ten MI patients with rEF and six MI patients with pEF, admitted to Intensive Cardiac Care after a first acute MI episode, were studied. The ECG was acquired during a Holter recording and the tachogram was extracted. Sample entropy (SE) and Lempel-Ziv Complexity (LZC 1 and LZC 2) metrics were computed on five hour long tachogram portions. A significant correlation was found between LZC indices and EF in the whole population; SE, LZC 1 and LZC 2 were significantly higher in patients with pEF. Our results indicate that lower complexity characterizes the HRV of MI patients with rEF. Complexity reduction might be due to a simplification of regulatory mechanisms, which might explain why MI patients with rEF are at higher risk for subsequent non-fatal and fatal events.

Prediction of acute hypotensive episodes by means of neural network multi-models.

This work proposes the application of neural network multi-models to the prediction of adverse acute hypotensive episodes (AHE) occurring in intensive care units (ICU). A generic methodology consisting of two phases is considered. In the first phase, a correlation analysis between the current blood pressure time signal and a collection of historical blood pressure templates is carried out. From this procedure the most similar signals are determined and the respective prediction neural models, previously trained, selected. Then, in a second phase, the multi-model structure is employed to predict the future evolution of current blood pressure signal, enabling to detect the occurrence of an AHE. The effectiveness of the methodology was validated in the context of the 10th PhysioNet/Computers in Cardiology Challenge-Predicting Acute Hypotensive Episodes, applied to a specific set of blood pressure signals, available in MIMIC-II database. A correct prediction of 10 out of 10 AHE for event 1 and of 37 out of 40 AHE for event 2 was achieved, corresponding to the best results of all entries in the two events of the challenge. The generalization capabilities of the strategy was confirmed by applying it to an extended dataset of blood pressure signals, also collected from the MIMIC-II database. A total of 2344 examples, selected from 311 blood pressure signals were tested, enabling to obtain a global sensitivity of 82.8% and a global specificity of 78.4%.