From Free Text to Structured Sections: A RapidMiner Workflow and Method Comparison for Slovak Medical Reports

Clinical documentation in Slovak hospitals is largely stored as free-running narrative text, which limits its reuse for decision support and secondary analysis. Slovak is a morphologically rich, comparatively low-resourced language, so general-purpose text-mining tools transfer poorly. This paper describes a complete, reproducible pipeline for segmenting Slovak medical reports into clinically meaningful sections and categorising the resulting paragraphs. Reports are split into paragraphs with a purpose-built Apache POI tool, normalised through the Morphonary lemmatiser extended with a custom abbreviation dictionary, and pre-processed in RapidMiner (tokenisation, stop-word removal, token-length filtering, term-frequency vectors). On a corpus of 232 paragraphs from 25 internal-medicine reports, we compare unsupervised clustering (k-means, k-medoids) against supervised classification (Naive Bayes, k-nearest neighbours, decision tree) under identical pre-processing and cosine similarity. Clustering recovered only partial, inconsistent structure, whereas supervised classification was far more effective: Naive Bayes was both the most accurate (89.5% paragraph-level accuracy) and the most balanced across classes, ahead of k-NN (87.5%) and a decision tree (72.4%), and required no parameter tuning. We report all operators and parameters for reproducibility, analyse per-class errors, and discuss the single-facility limitation. This work consolidates and extends our earlier published studies.