Improving Palliative Care with Deep Learning (1711.06402v1)

Abstract: Improving the quality of end-of-life care for hospitalized patients is a priority for healthcare organizations. Studies have shown that physicians tend to over-estimate prognoses, which in combination with treatment inertia results in a mismatch between patients wishes and actual care at the end of life. We describe a method to address this problem using Deep Learning and Electronic Health Record (EHR) data, which is currently being piloted, with Institutional Review Board approval, at an academic medical center. The EHR data of admitted patients are automatically evaluated by an algorithm, which brings patients who are likely to benefit from palliative care services to the attention of the Palliative Care team. The algorithm is a Deep Neural Network trained on the EHR data from previous years, to predict all-cause 3-12 month mortality of patients as a proxy for patients that could benefit from palliative care. Our predictions enable the Palliative Care team to take a proactive approach in reaching out to such patients, rather than relying on referrals from treating physicians, or conduct time consuming chart reviews of all patients. We also present a novel interpretation technique which we use to provide explanations of the model's predictions.

• The paper presents a deep neural network trained on over 2 million EHR records to accurately predict patient mortality within a 3-12 month window.
• It achieves an AUROC of 0.93 and a recall of 0.34 at a precision threshold of 0.9, outperforming traditional prognostication methods.
• The study incorporates an explanation technique to boost clinician trust by revealing key factors behind the model's predictions.

Improving Palliative Care with Deep Learning: An Expert Overview

The paper "Improving Palliative Care with Deep Learning" introduces a novel approach to addressing a significant challenge in healthcare: the timely identification of patients who could benefit from palliative care. Despite the increasing availability of palliative care services, a substantial number of patients who would benefit from such care do not receive it. This can be attributed in part to physician overestimation of patient prognoses, limited access due to resource constraints, and the inefficiencies of manual chart reviews. The authors propose utilizing deep learning algorithms, specifically deep neural networks, trained on Electronic Health Record (EHR) data to predict patient mortality within a 3-12 month window. This predictive capability can then facilitate targeted interventions by the palliative care team.

Methodology and Design

The core methodology employed by the authors involves the development of a deep learning model trained on historical EHR data from the STRIDE database, encompassing over 2 million patient records. The model is designed to perform as a binary classification system predicting the likelihood of mortality within a year of a specific date. Notably, the model is disease-agnostic, considering a wide array of medical codes and patient data, which include demographics and visit histories. The authors highlight the importance of a well-formulated selection criterion for training positive and negative cases to ensure the reliability of their proxy problem of mortality prediction.

The resulting deep neural network consists of an 18-layer architecture optimized using the SeLU activation function and Adam optimization. The authors placed emphasis on precise model training and evaluation, employing the Average Precision Score due to the class imbalance inherent in the dataset.

Results and Implications

Empirical results indicate that the predictive model exhibits strong performance, achieving an AUROC of 0.93 and a recall of 0.34 at a precision threshold of 0.9. The model is particularly notable for providing objective recommendations that counteract the potential biases of human physicians in prognostication. By freeing palliative care teams from exhaustive chart reviews, the system promises to make palliative care outreach more efficient and directed.

A significant contribution of this paper is the development of an explanation technique intended to elucidate why particular predictions were made. Given the inherent complexity of deep learning models, this approach aims to boost clinician trust in the model's predictions by identifying the key contributing factors for each patient’s mortality risk.

Practical and Theoretical Implications

From a practical viewpoint, this research showcases the ability of advanced machine learning models to handle complex medical datasets, providing highly impactful clinical insights. The pilot integration of this system in daily palliative care practice could lead to significant improvements in patient outcomes by enabling more timely and appropriate care planning.

Theoretically, the paper underscores the potential of deep learning to surpass traditional clinical scoring systems by leveraging large EHR datasets. While existing prognostic tools are often limited by narrow scope applications or require significant manual input, deep learning models can provide robust, scalable solutions across diverse patient populations without the necessity for disease-specific tailoring.

Speculation on Future Developments

The ongoing advancements in AI and machine learning promise further refinements to predictive models in healthcare. Future research could focus on enhancing model interpretability, exploring alternative proxy problems beyond mortality prediction, and integrating real-time data analysis. Moreover, interdisciplinary collaboration between AI researchers, clinicians, and policy makers will be essential to navigate ethical considerations and ensure equitable access to such technologies. As healthcare systems continue to digitize, leveraging machine learning will likely become integral to many aspects of patient care, from diagnostic support to treatment personalization.

This paper makes a strong case for the potential of EHR-driven predictions and sets a precedent for similar applications in other areas of medicine.