revisiting deep learning models for tabular data

#REVISITING DEEP LEARNING MODELS FOR TABULAR DATA UPDATE#

the splitting decision) thereby limiting their applicability 5. On the other hand, using FL to train machine learning models based on decision trees requires costly data transformations (e.g., homomorphic encryption) to share their sensitive information (e.g. FL is easily applicable when training neural networks and sharing model parameters, such as weights and bias. The role of the central service is key, as it orchestrates the distributed learning process. This can be accomplished by a central service that behaves as a coordinator, and in turn aggregates and sends back the model updates to the local sites. In a clinical scenario, FL allows to build a single model by distributing the training process across local sites (research centers and/or hospitals), thus avoiding the need to transfer data to a centralized server, and preserving data privacy. FL has been already exploited in the field of medical imaging for organs and tumor segmentations, as well as in the tabular data domain to predict hospitalization, mortality, length-of-stay in the intensive care unit (ICU) and outcomes in Covid-19 patients 3, 4. By capturing data variability and analyzing different demographics, FL has the potential to train distributed models that generalize across the entire artificial intelligence (AI) health domain 2. This leads to models that are trained on limited data and may perform poorly on other datasets 2.

This means that data are processed in a centralized manner, where both data collection and modelling are locally performed. One of the reasons is that data transactions are hampered by data access and privacy constrains, limiting the data to form data islands 1. In summary, we demonstrate the feasibility of a scalable FL for detecting hemodynamic instability in ICU data, which preserves privacy and holds the deployment benefits of a neural-based architecture.įederated learning (FL) is gaining momentum in data-driven innovation. Comparable performances in terms of accuracy, precision, recall and F1 score have been reached when detecting hemodynamic instability in FL, and in a centralized setup. This enhancement enables decision tree models in FL, which represent the state-of-the-art approach for classification tasks involving tabular clinical data. We employ a clinical dataset from 25 hospitals generated based on the Philips eICU database and we design a FL pipeline that supports neural-based boosting models as well as conventional neural networks. In this work, we focus on predicting hemodynamic instability on ICU patients by enabling distributed gradient boosting in FL. The latter, when embedded in FL, suffer from costly encryption techniques applied for sharing sensitive information such as the splitting decisions within the trees. Although widely applied on neural networks, the deployment of FL architectures is lacking scalability and support for machine learning techniques such as decision tree-based models.

#REVISITING DEEP LEARNING MODELS FOR TABULAR DATA UPDATE#

To do so, after having sent model updates to a central server, an update aggregation is performed, and the model is sent back to the sites for further training. FL enables hospitals to collaboratively learn a shared prediction model without moving the data outside their secure infrastructure. Federated learning (FL) is a privacy preserving approach to learning that overcome issues related to data access, privacy, and security, which represent key challenges in the healthcare sector.

Revisiting deep learning models for tabular data