Electronic Health Records (EHRs) have revolutionized the way patient data is collected and stored in healthcare settings. EHRs contain a wealth of patient data, such as patient demographics, medical history, laboratory results, clinical notes, and other information, which provides healthcare professionals with more comprehensive information about patients. The vast amounts of patient information stored in EHRs can enable clinicians to make more informed decisions about patient care [1]. EHRs have also enabled data-driven predictions of drug effects and interactions, patient diagnosis, and other applications [2], paving the way for precision medicine and enabling tailored treatments based on individual patient characteristics, leading to improved patient outcomes.

Machine learning models are capable of learning patterns and relationships in EHR data to assist clinicians and healthcare professionals in making more informed decisions about patient care. Clinical machine learning often relies on feature selection and data representation techniques to identify the most relevant features for each clinical application. However, manual feature selection can be time-consuming and labor-intensive, making it difficult to identify the relevant features for each clinical application, particularly given the large size of EHR datasets[3]. Additionally, this approach may reduce the generalizability of the framework, as the selected features can only be applied to that particular dataset or application, limiting the potential for transferability and scalability[4]. Another drawback of manual feature selection is that it may lead to the exclusion of important clinical features that could have high predictive power. Consequently, specialized algorithms have been proposed to effectively represent EHR data, making them more effective in analyzing and predicting clinical outcomes across different patient populations and applications [5], [6].

Representation learning is a type of machine learning that learns the fundamental correlations among the data points and represents them in a lower-dimensional space[7]. Deep patient representation learning utilizes deep learning algorithms to automatically extract and learn features from large and complex EHR data. The resulting representation effectively encapsulates significant clinical information, comprising medical history, vital signs, medication records, and laboratory findings. This representation can be used for various downstream healthcare applications, such as patient stratification[5], disease diagnosis [8], and drug discovery [9]. Numerous deep patient representation learning models have been proposed in the literature that utilizes machine learning algorithms like autoencoders[10], Recurrent Neural Networks (RNN)[11], and graph neural networks[12]. While these models have the potential to offer significant benefits in improving patient outcomes, they may suffer from biases that can lead to negative healthcare outcomes[13]. For instance, if the training data mostly comprises patients from a particular demographic group, such as Caucasians, the resulting representations may not generalize well to patients from other groups like African Americans or Asians. Additionally, if the training data does not represent the actual population distribution, the model may learn spurious correlations that do not reflect the underlying causal relationships between clinical features and outcomes[14]. Thus, it is essential to investigate the bias and fairness issues in deep patient representations and propose new models that can create unbiased patient representations.

In this paper, we investigate the presence of gender bias in deep patient representation learning models, and propose a novel unbiased patient representation model that utilizes an autoencoder architecture. The proposed “Fair Patient Model(FPM)” aims to generate unbiased and generalized patient representations that can be utilized for multiple downstream applications like patient mortality prediction, patient stratification, etc. FPM employs a customized loss function called the weighted reconstruction loss, which calculates the mean squared error loss for each patient and weights it by the inverse of its class frequency. To harness the collective strength of both structured and unstructured data, we integrate both these data modalities and apply FPM to derive more holistic patient representations, ensuring the fairness of the learned representations. We evaluate the model using the MIMIC-III dataset to predict patient mortality across different subpopulations, and demonstrate that the FPM representations outperform both deep patient representation models and common debiasing methods in terms of fairness scores.