Methods Inf Med 2023; 62(01/02): 060-070
DOI: 10.1055/s-0043-1762904
Original Article

Evaluation of Feature Selection Methods for Preserving Machine Learning Performance in the Presence of Temporal Dataset Shift in Clinical Medicine

Joshua Lemmon
1   Program in Child Health Evaluative Sciences, The Hospital for Sick Children, Toronto, Ontario, Canada
Lin Lawrence Guo
1   Program in Child Health Evaluative Sciences, The Hospital for Sick Children, Toronto, Ontario, Canada
Jose Posada
2   Biomedical Informatics Research, Stanford University, Palo Alto, California, United States
3   Department of Systems Engineering, Universidad del Norte, Barranquilla, Atlantico, Colombia
Stephen R. Pfohl
2   Biomedical Informatics Research, Stanford University, Palo Alto, California, United States
Jason Fries
2   Biomedical Informatics Research, Stanford University, Palo Alto, California, United States
Scott Lanyon Fleming
2   Biomedical Informatics Research, Stanford University, Palo Alto, California, United States
Catherine Aftandilian
4   Division of Pediatric Hematology/Oncology, Stanford University, Palo Alto, California, United States
Nigam Shah
2   Biomedical Informatics Research, Stanford University, Palo Alto, California, United States
Lillian Sung
1   Program in Child Health Evaluative Sciences, The Hospital for Sick Children, Toronto, Ontario, Canada
5   Division of Haematology/Oncology, The Hospital for Sick Children, Toronto, Ontario, Canada
› Author Affiliations
Funding L.S. is supported by the Canada Research Chair in Pediatric Oncology Supportive Care.


Background Temporal dataset shift can cause degradation in model performance as discrepancies between training and deployment data grow over time. The primary objective was to determine whether parsimonious models produced by specific feature selection methods are more robust to temporal dataset shift as measured by out-of-distribution (OOD) performance, while maintaining in-distribution (ID) performance.

Methods Our dataset consisted of intensive care unit patients from MIMIC-IV categorized by year groups (2008–2010, 2011–2013, 2014–2016, and 2017–2019). We trained baseline models using L2-regularized logistic regression on 2008–2010 to predict in-hospital mortality, long length of stay (LOS), sepsis, and invasive ventilation in all year groups. We evaluated three feature selection methods: L1-regularized logistic regression (L1), Remove and Retrain (ROAR), and causal feature selection. We assessed whether a feature selection method could maintain ID performance (2008–2010) and improve OOD performance (2017–2019). We also assessed whether parsimonious models retrained on OOD data performed as well as oracle models trained on all features in the OOD year group.

Results The baseline model showed significantly worse OOD performance with the long LOS and sepsis tasks when compared with the ID performance. L1 and ROAR retained 3.7 to 12.6% of all features, whereas causal feature selection generally retained fewer features. Models produced by L1 and ROAR exhibited similar ID and OOD performance as the baseline models. The retraining of these models on 2017–2019 data using features selected from training on 2008–2010 data generally reached parity with oracle models trained directly on 2017–2019 data using all available features. Causal feature selection led to heterogeneous results with the superset maintaining ID performance while improving OOD calibration only on the long LOS task.

Conclusions While model retraining can mitigate the impact of temporal dataset shift on parsimonious models produced by L1 and ROAR, new methods are required to proactively improve temporal robustness.

Consent for Publication

Not applicable.

Data Availability Statement

The data that support the findings of this study are available from PhysioNet but restrictions apply to the availability of these data, which were used under license for the current study, and thus not publicly available. Data are however available from the authors upon reasonable request and with permission of PhysioNet.

Ethical Approval Statement

The institutional review boards of Beth Israel Deaconess Medical Center, Boston, Massachusetts and the Massachusetts Institute of Technology, Cambridge, Massachusetts, United States waived the need for ethics approval and consequently participant informed consent due to the deidentification of patient records. All methods were performed in accordance with relevant guidelines and regulations.

Authors' Contributions

L.L.G. and L.S. designed the project with input from all authors. J.P. suggested the use of causal inference models. J.L. performed all experiments. J.L., L.L.G., and L.S. analyzed and interpreted results, with some input from all other authors. J.L. wrote the manuscript with major contributions from L.L.G. and L.S. All authors revised and commented on the manuscript. All authors read and approved the final manuscript.

Supplementary Material

Publication History

Received: 02 September 2022

Accepted: 04 January 2023

Article published online:
22 February 2023

© 2023. Thieme. All rights reserved.

Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany

  • References

  • 1 Moreno-Torres JG, Raeder T, Alaiz-Rodríguez R, Chawla NV, Herrera F. A unifying view on dataset shift in classification. Pattern Recognit 2012; 45 (01) 521-530
  • 2 Guo LL, Pfohl SR, Fries J. et al. Systematic review of approaches to preserve machine learning performance in the presence of temporal dataset shift in clinical medicine. Appl Clin Inform 2021; 12 (04) 808-815
  • 3 Davis SE, Greevy Jr RA, Fonnesbeck C, Lasko TA, Walsh CG, Matheny ME. A nonparametric updating method to correct clinical prediction model drift. J Am Med Inform Assoc 2019; 26 (12) 1448-1457
  • 4 Siregar S, Nieboer D, Versteegh MIM, Steyerberg EW, Takkenberg JJM. Methods for updating a risk prediction model for cardiac surgery: a statistical primer. Interact Cardiovasc Thorac Surg 2019; 28 (03) 333-338
  • 5 Chandrashekar G, Sahin F. A survey on feature selection methods. Comput Electr Eng 2014; 40 (01) 16-28
  • 6 Johnson A, Bulgarelli L, Pollard T, Horng S, Celi LA, Mark R. MIMIC-IV (version 1.0). PhysioNet. 2021. Accessed January 27, 2023 at:
  • 7 Goldberger AL, Amaral LA, Glass L. et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 2000; 101 (23) E215-E220
  • 8 Singer M, Deutschman CS, Seymour CW. et al. The third international consensus definitions for sepsis and septic shock (Sepsis-3). JAMA 2016; 315 (08) 801-810
  • 9 Tibshirani R. Regression shrinkage and selection via the lasso. J R Stat Soc B 1996; 58 (01) 267-288
  • 10 Hooker S, Erhan D, Kindermans P-J, Kim B. A benchmark for interpretability methods in deep neural networks. arXiv preprint arXiv:180610758. 2018. Accessed January 27, 2023 at:
  • 11 Yu K, Guo X, Liu L. et al. Causality-based feature selection: methods and evaluations. ACM Comput Surv 2020; 53 (05) 1-36 (CSUR)
  • 12 Tsamardinos I, Aliferis CF. Towards principled feature selection: relevancy, filters and wrappers. Paper presented at: Proceedings of the Ninth International Workshop on Artificial Intelligence and Statistics; Key West, Florida, United States, January 3–6, 2003; Proceedings of Machine Learning Research
  • 13 Tsamardinos I, Aliferis CF, Statnikov AR, Statnikov E. Algorithms for large scale Markov blanket discovery. Paper presented at: FLAIRS Conference; St. Augustine, Florida, United States, May 12–14, 2003
  • 14 Pena JM, Nilsson R, Björkegren J, Tegnér J. Towards scalable and data efficient learning of Markov boundaries. Int J Approx Reason 2007; 45 (02) 211-232
  • 15 De Morais SR, Aussem A. A novel scalable and data efficient feature subset selection algorithm. Paper presented at: Joint European Conference on Machine Learning and Knowledge Discovery in Databases; Antwerp, Belgium, September 15–19, 2008
  • 16 Aliferis CF, Statnikov A, Tsamardinos I, Mani S, Koutsoukos XD. Local causal and Markov blanket induction for causal discovery and feature selection for classification part I: algorithms and empirical evaluation. J Mach Learn Res 2010; 11 (07) 171-234
  • 17 Hassan A, Paik JH, Khare S, Hassan SA. PPFS: predictive permutation feature selection. arXiv preprint arXiv:211010713. 2021. Accessed January 27, 2023 at:
  • 18 Austin PC, Steyerberg EW. The Integrated Calibration Index (ICI) and related metrics for quantifying the calibration of logistic regression models. Stat Med 2019; 38 (21) 4051-4065
  • 19 Debray TP, Vergouwe Y, Koffijberg H, Nieboer D, Steyerberg EW, Moons KG. A new framework to enhance the interpretation of external validation studies of clinical prediction models. J Clin Epidemiol 2015; 68 (03) 279-289
  • 20 Guo LL, Pfohl SR, Fries J. et al. Evaluation of domain generalization and adaptation on improving model robustness to temporal dataset shift in clinical medicine. Sci Rep 2022; 12 (01) 2726
  • 21 Guo LL, Steinberg E, Fleming SL. et al. EHR foundation models improve robustness in the presence of temporal distribution shift. medRxiv 2022. Accessed January 27, 2023 at:
  • 22 Cawley GC. Causal & non-causal feature selection for ridge regression. Paper presented at: Causation and Prediction Challenge; Hong Kong, June 1–6, 2008
  • 23 Zhang X, Hu Y, Xie K, Wang S, Ngai E, Liu M. A causal feature selection algorithm for stock prediction modeling. Neurocomputing 2014; 142: 48-59