Naslov
The Disparate Goals of Statistics and Machine Learning: Survival Analysis and Prediction on Liver Transplantation Data

Autori
Stresec, Ivan ; Bezjak, Miran ; Jadrijević, Stipislav ; Kocman, Branislav ; Filipec Kanižaj, Tajana ; Mikulić, Danko ; Dalbelo Bašić, Bojana

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, ostalo

Izvornik
Book of Abstracts from the 26th International Scientific Symposium on Biometrics / Jazbec, Anamarija ; Tafro, Azra ; Šimić, Diana ; Pecina , Marija ; Vedriš, Mislav ; Sović, Slavica ; Brajković, Vladimir ; Sonicki, Zdenko - Zagreb : Hrvatsko biometrijsko društvo, 2023, 27-27

Skup
26th International Scientific Symposium on Biometrics (BIOSTAT2023)

Mjesto i datum
Zadar, Hrvatska, 14.06.2023. - 17.06.2023

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Nije recenziran

Ključne riječi
statistics ; machine learning ; survival analysis ; survival prediction ; liver transplantation

Sažetak
Survival analysis is a widely known branch of statistics that analyzes the duration of time until some event occurs. Regression modelling of survival data is somewhat atypical and predicated on the common occurrence of censored data. Like other biomedical data, survival data are often scarce due to the nature of their acquisition. In the last couple of decades, survival prediction, a machine learning (ML) approach to survival data, is becoming increasingly prominent. Like other supervised ML models, survival prediction models are designed to extract generalizable patterns from the data to make predictions on new, unseen data. The performance of an ML survival model is most commonly measured by the concordance index (c-index), a metric that tells us how well a model orders samples based on their survival time. The difference in statistical and machine learning models is not quantifiable, making the comparison between the two nontrivial. Nevertheless, it is not uncommon to see nonsensical c-index comparisons of ill-developed Cox models and ML models. This nonperception of the disparity between traditional statistical models and machine learning might indicate a lack of understanding on the practitioners' side. Our current research is concerned with both analysis and prediction of survival times of liver transplantation patients. While analyzing survival times in terms of donor and recipient variables can be used to infer knowledge about a sample to verify and challenge old and new hypotheses, survival prediction is more appropriate for systems designed to aid in donor-recipient matching decisions. We discuss our analysis of the survival outcomes of patients who underwent liver transplantation for hepatocellular carcinoma (HCC). Conversely, we also mention our performance-based comparison between statistical prognostic scores and commonly used survival ML models in the context of donor-recipient matching. As a final point of interest, we briefly touch upon interpretation. The interpretation of ML models is inherently more complicated and less transparent than the interpretation of statistical models. The comparison between the models' interpretations, much like the comparison between the models themselves, is not straightforward.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo, Biotehnologija u biomedicini (prirodno područje, biomedicina i zdravstvo, biotehničko područje)

POVEZANOST RADA