Paul Youssef

Paul Youssef

Auto-generated profile, based on publicly available information

Paul Youssef

Mathematics and Computer Science, Marburg University, Marburg, Germany

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
1 August 2024 in JCO Clinical Cancer Informatics
Verified
Journal Article
Closed access
MCMaria Rosa Salvador Comino
Maria Rosa Salvador Comino
PYPaul Youssef
Paul Youssef
AHAnna Heinzelmann
Anna Heinzelmann
FBFlorian Bernhardt
Florian Bernhardt
CSChristin Seifert
Christin Seifert
MTMitra Tewes
Mitra Tewes
PURPOSE: Palliative care is recommended for patients with cancer with a life expectancy of <12 months. Machine learning (ML) techniques can help in predicting survival outcomes among patients with cancer and may help distinguish who benefits the most from palliative care support. We aim to explore the importance of several objective and subjective self-reported variables. Subjective variables were collected through electronic psycho-oncologic and palliative care self-assessment screenings. We used these variables to predict 1-year mortality. MATERIALS AND METHODS: Between April 1, 2020, and March 31, 2021, a total of 265 patients with advanced cancer completed a patient-reported outcome tool. We documented objective and subjective variables collected from electronic health records, self-reported subjective variables, and all clinical variables combined. We used logistic regression (LR), 20-fold cross-validation, decision trees, and random forests to predict 1-year mortality. We analyzed the receiver operating characteristic (ROC) curve—AUC, the precision-recall curve—AUC (PR-AUC)—and the feature importance of the ML models. RESULTS: The performance of clinical nonpatient variables in predictions (LR reaches 0.81 [ROC-AUC] and 0.72 [F1 score]) are much more predictive than that of subjective patient-reported variables (LR reaches 0.55 [ROC-AUC] and 0.52 [F1 score]). CONCLUSION: The results show that objective variables used in this study are much more predictive than subjective patient-reported variables, which measure subjective burden. These findings indicate that subjective burden cannot be reliably used to predict survival. Further research is needed to clarify the role of self-reported patient burden and mortality prediction using ML.

8 August 2026
Verified
Conference Paper
Open access
YCYinjie Cheng
Yinjie Cheng
PYPaul Youssef
Paul Youssef
CSChristin Seifert
Christin Seifert
JSJörg Schlötterer
Jörg Schlötterer
ZZZhixue Zhao
Zhixue Zhao
Knowledge editing (KE) offers a lightweight alternative to retraining for updating large language models (LLMs). Meanwhile, fine-tuning remains the default operation for adapting LLMs to new domains and tasks. Despite their widespread adoption, these two post-training interventions have been studied in isolation, leaving open a crucial question: if we fine-tune an edited model, do the edits survive? This question is motivated by practical objectives: removing covert or malicious edits, and preserving beneficial edits. If fine-tuning impairs edits (Fig.1), current KE methods become less efficient, as a newly fine-tuned model requires re-editing; if edits persist, fine-tuned models risk propagating hidden malicious edits, raising serious safety concerns. To this end, we systematically quantify edit decay after fine-tuning across 254 experimental configurations. Our results show that in general, edits decay substantially after subsequent fine-tuning. AlphaEdit exhibits the greatest decay on the zsRE benchmark when applied to GPT-J, where 25.27% of previously successful edits become unsuccessful after fine-tuning. We further find that fine-tuning only the edited layers is sufficient to effectively remove edits, while incurring only modest degradation in downstream performance. Surprisingly, fine-tuning non-edited layers leads to greater edit decay than all-layer fine-tuning. Besides, our activation space analysis reveals that fine-tuning produces a larger and more coherent representational shift, both in magnitude and direction, than KE. Overall, our study underscores the necessity of evaluating KE within the broader LLM application pipeline. Our code and results are available at https://github.com/Cheng-Yinjie/edit_decay_KDD.