% IMPORTANT: The following is UTF-8 encoded.  This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.

@ARTICLE{Nieto:1051639,
      author       = {Nieto, Nicolas and Raimondo, Federico and Eickhoff, Simon
                      and Poess, Janine and Desch, Steffen and Feistritzer,
                      Hans-Josef and Lichtenberg, Artur and Masyuk, Maryna and
                      Kelm, Malte and Thiele, Holger and Jung, Christian and
                      Patil, Kaustubh},
      title        = {{M}achine learning models for early prognosis prediction in
                      cardiogenic shock},
      journal      = {MedRxiv},
      reportid     = {FZJ-2026-00553},
      year         = {2025},
      abstract     = {Cardiogenic shock (CS) is a severe and frequent
                      complication of acute myocardial infarction (AMI),
                      necessitating rapid and accurate prognosis as-sessment to
                      guide treatment and intensive care unit (ICU) resource
                      allocation. We developed two machine learning models to
                      predict 30-day outcomes following CS in AMI: an Admission
                      model (using only data available at admission, like
                      demography, comorbidities) and a Full model (incorporating
                      additional laboratory values obtained within 24 hours). The
                      models were trained on the CULPRIT-SHOCK dataset and
                      externally validated using the eICU database. The Admission
                      model achieved an out-of-sample AUC of 0.71 $(95\%$ CI:
                      0.6–0.83) in the development cohort and 0.68 in the
                      validation cohort, while the Full model attained
                      significantly higher performance, with AUCs of 0.80 $(95\%$
                      CI: 0.69–0.9) and 0.78, respectively. The Full model’s
                      superior performance underscores the prognostic value of
                      early laboratory trends, suggesting that dynamic data
                      integration improves risk stratification. Both models
                      outperformed existing risk scores across multiple metrics,
                      provided well-calibrated probabilistic predictions, and
                      demonstrated robustness to missing data. Additionally, they
                      offered patient-level explainability, enhancing clinical
                      interpretability. While promising, the models’
                      generalizability may be influenced by differences between
                      the CULPRIT-SHOCK and eICU cohorts (e.g., demographics, CS
                      severity thresholds); further validation in larger,
                      prospective cohorts is warranted.},
      cin          = {INM-7},
      cid          = {I:(DE-Juel1)INM-7-20090406},
      pnm          = {5254 - Neuroscientific Data Analytics and AI (POF4-525)},
      pid          = {G:(DE-HGF)POF4-5254},
      typ          = {PUB:(DE-HGF)25},
      doi          = {10.1101/2025.09.18.25336054},
      url          = {https://juser.fz-juelich.de/record/1051639},
}