Healthcare data, particularly in critical care settings, presents three key challenges for analysis. First, physiological measurements come from different sources but are inherently related. Yet, traditional methods often treat each measurement type independently, losing valuable information about their relationships. Second, medical measurements are collected at irregular intervals, and these sampling times can carry clinical meaning. Finally, whilst several imputation methods exist to tackle the common problem of missing values, they often fail to address the temporal nature of the data or provide estimates of uncertainty in their predictions. We propose using deep Gaussian process emulation using stochastic imputation, a methodology initially conceived to deal with computationally expensive calculation for uncertainty quantification, to solve the problem of handling missing values that naturally occur in critical care data. This method leverages longitudinal and cross-sectional information and provides uncertainty estimation for the imputed values. Our evaluation of a clinical dataset shows that the proposed method performs better than conventional methods, such as multiple imputations with chained equations (MICE), last-known value imputation, and individually fitted Gaussian Processes (GPs).

This pattern is evident in several studies across various aspects of intensive care. A recent study involving over 500,000 participants revealed that women were less likely than men to receive mechanical ventilation or renal replacement therapy and had shorter ICU stays (Modra et al., 2022). Similarly, Todorov et al. (2021), in a study of comparable size, also observed that female cardiovascular patients were less likely to receive intensive care than men despite being more severely ill. Pietropaoli et al. (2010) noted differences in ICU admission rates for severe sepsis or septic shock between genders, though they concluded that the higher in-hospital mortality risk for women was not associated with differences in care delivery. Finally, using MIMIC and eICU, two publicly available intensive care datasets, Liu et al. (2022) discovered that women are subject to inequality, such as treatment delays, lower rates of ventilation, and a higher likelihood of choosing limited care.

Unfortunately, most of the studies mentioned above did not identify the potential sources of bias, as establishing causal relationships requires comprehensive analysis, often involving counterfactuals that are difficult or sometimes impossible to obtain in clinical settings. However, as Pietropaoli et al. (2010) also summarised, gender bias in code status on ICU admission may arise from two potential sources: women (or their surrogates) tending to prefer less aggressive treatments (as shown by Bookwala et al. (2021) and Wenger (1995)), or healthcare providers influencing end-of-life decisions while misunderstanding patients’ actual preferences (as shown by Cook (1995)).

As plausible steps to prevent gender bias, Todorov et al. (2021) suggested that intensivists and emergency physicians reassess triage decisions to ensure critically ill (young) women receive necessary care, addressing potential gender biases in protocols. Research into gender-specific factors could also lead to targeted therapies (Pietropaoli et al., 2010). Additionally, this effort should not be separated from educational initiatives to identify and address gender biases.

Bibliography

Block, L. et al. (2019) ‘Age, SAPS 3 and female sex are associated with decisions to withdraw or withhold intensive care’, Acta Anaesthesiologica Scandinavica, 63(9), pp. 1210–1215. Available at: https://doi.org/10.1111/aas.13411.

Bookwala, Kristen M. Coppola, Angel, J. (2001) ‘GENDER DIFFERENCES IN OLDER ADULTS’ PREFERENCES FOR LIFE-SUSTAINING MEDICAL TREATMENTS AND END-OF-LIFE VALUES’, Death Studies, 25(2), pp. 127–149. Available at: https://doi.org/10.1080/07481180126202.

Cook, D.J. (1995) ‘Determinants in Canadian Health Care Workers of the Decision to Withdraw Life Support From the Critically Ill’, JAMA: The Journal of the American Medical Association, 273(9), p. 703. Available at: https://doi.org/10.1001/jama.1995.03520330033033.

Liu, X. et al. (2022) ‘Evaluating Prognostic Bias of Critical Illness Severity Scores Based on Age, Gender, and Primary Language in the USA: A Retrospective Multicenter Study’. Available at: https://doi.org/10.1101/2022.08.01.22277736.

Modra, L.J. et al. (2022) ‘Sex Differences in Treatment of Adult Intensive Care Patients: A Systematic Review and Meta-Analysis’, Critical Care Medicine, 50(6), pp. 913–923. Available at: https://doi.org/10.1097/CCM.0000000000005469.

Todorov, A. et al. (2021) ‘Gender differences in the provision of intensive care: a Bayesian approach’, Intensive Care Medicine, 47(5), pp. 577–587. Available at: https://doi.org/10.1007/s00134-021-06393-3.

Pietropaoli, A.P. et al. (2010) ‘Gender differences in mortality in patients with severe sepsis or septic shock’, Gender Medicine, 7(5), pp. 422–437. Available at: https://doi.org/10.1016/j.genm.2010.09.005.

Wenger, N.S. (1995) ‘Epidemiology of Do-Not-Resuscitate Orders: Disparity by Age, Diagnosis, Gender, Race, and Functional Impairment’, Archives of Internal Medicine, 155(19), p. 2056. Available at: https://doi.org/10.1001/archinte.1995.00430190042006.

@article{septiandri2024bias,
  title   = {Gender Bias in Intensive Care},
  author  = {Septiandri, Ali Akbar},
  year    = 2024,
  month   = {October},
  url     = {https://aliakbars.id/posts/2024/10/gender-bias}
}

In statistical modeling, it is common to encounter missing values, particularly in healthcare situations where vital sign measurements may be sporadic. This can lead to difficulties in accurately evaluating a patient’s health and making informed decisions about their treatment. One possible solution is to use imputation methods to fill in the missing values and enable counterfactual predictions in healthcare settings. To address this issue, we have conducted a study examining existing approaches to missing value imputation in healthcare settings, with a specific focus on time series data. We provide an overview of both classical (e.g. MICE, MissForest, Gaussian Processes), and deep learning-based (e.g. Recurrent Neural Networks, Generative Adversarial Networks, Autoencoders) imputation methods, assess their strengths and limitations within the context of time series data, and highlight potential areas for further investigation.

@article{septiandri2023missing,
  title   = {Handling Missing Values in Healthcare Settings},
  author  = {Septiandri, Ali Akbar and Jendoubi, Takoua and DíazDelaO, F. Alejandro},
  year    = 2023,
  month   = {September},
  url     = {https://aliakbars.id/posts/2023/09/missing-values}
}