gms | German Medical Science

In order to develop individual treatment rules (ITR) for moderate Community-Acquired Pneumonia (CAP) we used different Machine Learning (ML) methods – such as decision trees or scoring models - based on aetiological and clinical parameters, which are available ex ante. The ITR - is an individual suggestion which kind of antibiotics the patient should be treated with. We computed an expected mortality for each patient by logistic regression and took the difference to the real mortality as numeric class-label for ML.

To consider seasonal effects we displayed characteristics of more than 9,000 hospitalised patients with moderate severity (non-intensive care unit patients) from the observational, prospective, multinational CAPNETZ study [1] by the treatment day within the year. We could observe interesting behaviours, e.g. low mortality of fluoroquinolones treated patients in spring, but firstly could not prove a significant relation to the Influenza seasons.

Each Influenza season in Germany has an individual distribution. Weekly case-numbers from 2001 onwards are publicly available from the RKI [2]. By normalization, we eliminated the fact that during the years more cases are reported. We also constructed daily case numbers by interpolation.

We then gave a patient of CAPNETZ a higher weight if the normalized Influenza case number of the treatment day is higher. This reflects the assumption that the probability of a patient in the database being infected by Influenza is related to the RKI cases at the treatment day. In addition we considered not only the treatment day itself but also a fixed time shift (e.g. exactly 10 days before treatment).

With this method (convolution) the significance of our observation improved considerably. In addition, we found that the lower mortality of fluoroquinolones treated patients goes along with an increase of the overall severity of CAP after the Influenza season.

The method presented here can be generalized to other data or other pathogens where case-numbers are available.

Members of the CAPNETZ study group are: M. Dreher, C. Cornelissen (Aachen); W. Knüppel (Bad Arolsen); D. Stolz (Basel); N. Suttorp, M. Witzenrath, P. Creutz, A. Mikolajewska (Berlin, Charité); T. Bauer, D. Krieger (Berlin); W. Pankow, D. Thiemig (Berlin-Neukölln); B. Hauptmeier, S. Ewig, D. Wehde (Bochum); M. Prediger, S. Schmager (Cottbus); M. Kolditz, B. Schulte-Hubbert, S. Langner (Dresden); W. Albrich (St Gallen); T. Welte, J. Freise, G. Barten, O. Arenas Toro, M. Nawrocki, J. Naim, M. Witte, W. Kröner, T. Illig, N. Klopp (Hannover); M. Kreuter, F. Herth, S. Hummler (Heidelberg); P. Ravn, A. Vestergaard-Jensen, G. Baunbaek-Knudsen (HillerØd); M. Pletz, C. Kroegel, J. Frosinski, J. Winning, B. Schleenvoigt (Jena); K. Dalhoff, J. Rupp, R. Hörster, D. Drömann (Lübeck); G. Rohde, J. Drijkoningen, D. Braeken (Maastricht); H. Buschmann (Paderborn); T. Schaberg, I. Hering (Rotenburg/Wümme); M. Panning (Freiburg); M. Wallner (Ulm).

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.