›› Introduction

The aim of machine learning is to discover the implicit structure or model of some data set. Among the many automatic learning learning methods that have been proposed, inductive logic programming (ILP) is particularly interesting because it enables a deep modeling of complex data: such data are structured by several kinds of relations that it is difficult to make automatically explicit by more classical methods such as those based on an attribute-value representation.

ILP is a supervised symbolic machine learning method: from a set of examples describing situations where the concept is present or absent, the ILP learning method induces a representation of the concept that generalizes the examples. This representation takes the form of a set of first-order clauses. This last property is generally required for representing simply and naturally temporal patterns such as those associated to cardiac disorders. As a result, the learned rules are expressed in a high-level language and can be easily understood by a human. This is a very important aspect in medicine for validating the results to domain experts.

More information on ILP can be found here.

›› Contributions

Our work focusses on temporal knowledge discovery from time series recorded by sensors. Precisely, learning examples are electrocardiogram fragments related to disorders and expressed symbolically as event sequences. Learning with ILP results in a set of clauses containing literals related to specific events present in the electrocardiogram and temporal constraints on their co-occurrence. The translation of such sets of clauses into the chronicle formalism for chronicle recognition is straightforward.

Monosource learning

The representation of examples and hypotheses should facilitate learning, particularly with respect to the implicit structure of learning data. To this aim, we have introduced a chain structure for representing examples and hypotheses. While avoiding as much as possible unnecessary bias, this representation facilitates the learning of the sequential structure of events and their generalization in learned models. Moreover, the search space can be reduced by early pruning the paths leading to solutions that would not respect the adequate sequential structure.

Multisource learning

When the data come from different sources, e.g. several ECG channels and pressure or respiratory channels, the size of the search space becomes so large that learning by ILP is impractical. Thus, an idea is to use a divide and conquer method. Learning is achieved in two steps: learning independently from each source data using a specific representation language that is simpler than the global language and, then, merging the partial results to obtain the global model. We have proposed an original method consisting in not merging directly the partial models but in using them to search efficiently the global hypothesis space.

Learning decision rules

Decision rule acquisition, either from experts or using data analysis techniques such as principal component analysis, may be long and tedious. We have assessed a machine learning technique, i.e. decision tree induction, from the results of which decision rules for the piloting module can be derived. To this end, each concurrent algorithm is executed in multiple contexts with different kinds and levels of clinical noise. The premise of the generated rules are made of several tests on context attributes and their conclusion gives the chosen algorithm and a parameter setting.