Complications after surgery significantly impact perioperative patient outcome. Accurate risk stratification can influence indication for surgery as well as postoperative treatment.
Machine learning (ML) describes a class of new computer-based data analysis methods, which relies on pattern recognition to generate general assumptions and predict the possibilities of predefined events.
ML is therefore excellently suited to predict complications after surgery.
An application for such a ML based alert algorithm (CASSANDRA) can be the preoperative risk stratification of patients with the intention to uncover previously unidentified patterns and therefore enable an automatic assessment. Additionally, if such a system is continuously fed information during an in-patient stay, real-time predictions of complications can be possible. At the moment we are in the development of ML based complication recognition mechanisms and are planning an implementation of real-time alert algorithms in the near future.
Natural language processing (NLP) refers to the interpretation of natural language with the help of machine leaning algorithms. Object and pattern recognition mechanisms are used to extract information out of written text. Due to the complexity and ambiguity of the human language, gathered information, especially from medical documents, is difficult to interpret. Data can be presented as numbers for laboratory values, standard results text elements such as ECG reports or complex texts such as psychiatric evaluations. Until now, the available technology was insufficient for reliable analyses.
In a first proof of concept, we evaluated 600 surgical reports of pancreas resections for the correct classification of the reconstruction (either pancreaticojejonostomy or pancreaticogastrostomy). We split the dataset 3 to 1 in a training and test dataset. Best results yieled a naive bayesian classefier with an accuracy of 92%.
We show, that an automated classification of surgical reports via natural language processing is possible with an acceptable accuracy.