Analysis of respiratory pressure–volume curves in intensive care medicine using inductive machine learning

Abstract

We present a case study of machine learning and data mining in intensive care medicine. In the study, we compared different methods of measuring pressure–volume curves in artificially ventilated patients suffering from the adult respiratory distress syndrome (ARDS). Our aim was to show that inductive machine learning can be used to gain insights into differences and similarities among these methods. We defined two tasks: the first one was to recognize the measurement method producing a given pressure–volume curve. This was defined as the task of classifying pressure–volume curves (the classes being the measurement methods). The second was to model the curves themselves, that is, to predict the volume given the pressure, the measurement method and the patient data. Clearly, this can be defined as a regression task. For these two tasks, we applied C5.0 and CUBIST, two inductive machine learning tools, respectively. Apart from medical findings regarding the characteristics of the measurement methods, we found some evidence showing the value of an abstract representation for classifying curves: normalization and high-level descriptors from curve fitting played a crucial role in obtaining reasonably accurate models. Another useful feature of algorithms for inductive machine learning is the possibility of incorporating background knowledge. In our study, the incorporation of patient data helped to improve regression results dramatically, which might open the door for the individual respiratory treatment of patients in the future.

Introduction

Artificial ventilation is one of the key therapies in intensive care medicine. In particular if the lung is ill, it is not unproblematic, and there is always the risk of damaging it by barotrauma or volutrauma, respectively. Therefore, it is important to observe the mechanical status of the artificially ventilated lung. The two quantities most crucial for assessing the pathophysiological mechanical condition of the respiratory system (lung and thorax) are the compliance C (the volume distensibility), and the air-flow resistance R (both will be formally defined in the following section). There are two fundamentally different approaches to measuring compliance C: two-point and multiple-point methods. Two-point methods obtain measurements at two different volumes, usually at the end of inspiration and at the end of expiration, whereas multiple-point methods obtain measurements during the whole breath. In order to find the respective advantages and disadvantages of various two- and multiple-point methods, the university hospitals of Aachen, Berlin, Bonn, Freiburg i. Br., Göttingen, Hamburg, Mannheim-Heidelberg and Munich are conducting a multi-center study in collaboration with the company Drägerwerk AG, Lübeck. Three multi-point methods (LOW-FLOW, SUPER-SYRINGE, SLICE) as well as two two-point methods (static compliance by automated single steps (SCASS) and PEEP-WAVE) are applied in randomized order to each patient participating in the study.

The central issue of the multi-center study is to discover systematic differences among the pressure–volume data generated by these five different measurement methods. Our aim was to show that data mining methods, and in particular methods from inductive machine learning, are generally well-suited to detect structural characteristics of measured pressure–volume curves. Another goal was to obtain interpretable results, such that physicians could inspect them and relate them to prior knowledge. This would advance the state of the art in the area, since at present pressure–volume curves are “just” analyzed visually: important parameters like the lower inflection point LIP (to be defined later) and the upper inflection point UIP, are determined upon visual inspection of the pressure–volume curves. Both the lower and upper inflection point are critical for setting the parameters of artificial ventilatory systems. So, the application of data mining and machine learning tools might not only lead to new insights about the similarities and differences among these measurement methods, but ultimately contribute to more objective decisions.

Many different approaches can be taken to perform data mining. Some data mining techniques originate from the field of machine learning. One advantage of the machine learning methods used here is that they can flexibly handle all sorts of descriptors. For instance, in our domain it is possible to include patient data and see whether there is a connection between properties of the patients and the shape of the pressure–volume curves. In the long run, this might open the door for the individual treatment of patients under artificial ventilation.

From a machine learning perspective, this paper is concerned with the analysis of curves by means of classification and regression techniques. In particular, the classification of curves is an interesting topic, because surprisingly few papers in the literature deal with this type of classification task. Our experience was that normalization and curve fitting techniques dramatically improve the classification results. Apart from these findings, our machine learning approach may be even more widely applicable, since it yields interpretable models of the similarities and differences among curves.

The organization of this paper parallels the one in Cios et al. [1]. Section 2 details the medical problem domain, Section 3 presents the datasets and Section 4 the preparation of the data. Section 5 focuses on the actual data mining step: it describes the data modeling techniques, the training and test procedures, and the results from a quantitative point of view. Subsequently, the discovered knowledge is evaluated in Section 6. Section 7 dicusses the usage and potential usage of the discovered knowledge, before a section touching upon related work. Finally, we conclude the paper and hint at possible directions of further research.