@techreport{TR-IC-PFG-26-23,
   number = {IC-PFG-26-23},
   author  =  {Isabella  Ribeiro  Rigue,  Luigi  Mello Rigato, Gabriel
                   Bianchin de Oliveira and Zanoni Dias},
   title   =   {{Cancer  Survival  Prediction  through  Multifactorial 
                   Analysis with Machine Learning}},
   month = {July},
   year = {2026},
   institution = {Institute of Computing, University of Campinas},
   note = {In English, 31 pages.
    \par\selectlanguage{english}\textbf{Abstract}
       Cancer  remains  a  pressing public health challenge in Brazil,
       demanding  data-driven  strategies to improve patient outcomes.
       The  objective  of this work is to develop and evaluate machine
       learning  and  deep  learning  models  to  predict  the overall
       survival  of  patients  with  cancer  at  the  beginning of the
       initial   treatment,   using   demographic,   diagnostic,   and  
       treatment-related  information  available  at  that  time. This
       study  uses  a  large-scale dataset from the Brazilian National
       Cancer  Institute  (INCA).  By  processing over 5.4 million raw
       longitudinal  records  through  a  PySpark  pipeline to resolve
       severe  data quality and temporal inconsistencies, we extracted
       a refined cohort of 73,459 high-quality samples. Our predictive
       modeling   explored   binary   classification  (random  forest, 
       XGBoost,  multi-layer  perceptron,  AutoGluon,  and ensembles),
       achieving  high discriminative power with ROC-AUC scores around
       0.88,  and  advanced  survival  analysis  (DeepSurv and XGBoost
       Survival) for time-to-event dynamics, where the DeepSurv neural
       network    reached    a    C-index   of   approximately   0.85.  
       Interpretability analyses revealed that the primary tumor site,
       metastasis  staging, patient age, and treatment history are the
       most  critical  predictors of mortality risk. Ultimately, these
       results   validate   the  feasibility  of  applying  artificial 
       intelligence    to   Brazilian   hospital   oncology   records,  
       demonstrating the potential of these models for future clinical
       decision-support  applications within the Unified Health System
       (SUS).
  }
}