Federating a Quantile Regression Model for Neonatal Care at Tel Aviv (Sourasky) Medical Center Using Rhino Health’s Federated Computing Platform

Apr 9, 2024
Lili Lau, Director of Product Marketing
Yuval Baror, Co-founder & CTO
Noy Maimon, Backend Software Engineer

In a joint effort at the forefront of neonatal healthcare innovation, the I-Medata AI Center at the Tel Aviv (Sourasky) Medical Center has partnered with Rhino Health to assess and validate the capabilities of Federated Learning technology in improving machine learning (ML) algorithms for neonatal care outcomes’ prediction with Rhino Health providing the Federated Computing infrastructure essential for the project. Renowned for its pivotal work in AI application in medicine, the I-Medata AI Center, in collaboration with the Department of Neonatology at the Tel Aviv (Sourasky) Medical Center’s Dana-Dwek Children’s Hospital, developed a Quantile Regression AI model tailored for predicting the Length of Stay (LOS) for newborns in neonatal care. Combining the I-Medata AI Center expertise with Rhino Health’s advanced Federated Computing Platform (FCP), both parties embarked on a proof of concept. This collaboration aimed to rigorously protect patient data privacy while exploring the model’s potential in predicting the Length of Stay (LOS) for newborns in neonatal care.

"Rhino Health's Federated Computing Platform provided us with the secure and robust infrastructure needed to test our Quantile Regression AI model in a Federated way. With their support, we were able to simulate the development of predictive models for the care of our premature newborn patients based on data from different hospitals. Rhino Health's expertise in Federated Edge and Federated Learning streamlined our workflow and fortified the privacy and security of our patient data, which is paramount in our field. We look forward to continuing this fruitful partnership as we explore new frontiers in healthcare AI."
Brenda Kasabe, Data Engineering Project Management, I-Medata AI Center at Tel Aviv (Sourasky) Medical Center

1. The Need for Predictive Models in Neonatal Care

The prediction of LOS for premature newborns in neonatal intensive care units (NICUs) is a critical component of neonatal care. Accurate forecasts of LOS are valuable for several reasons: they enable healthcare providers to optimize resource allocation, enhance the planning and delivery of care, and, ultimately, improve patient outcomes. Preterm neonates, due to their vulnerability and the complex care they require, often necessitate a multidisciplinary approach that can significantly benefit from predictive insights into the duration of their hospital stay. Such predictions not only facilitate better management of NICU capacities but also assist in the timely identification of potential complications, thereby allowing for preemptive intervention strategies that can mitigate risks to these susceptible patients.

Delivering high-quality neonatal care is fraught with challenges, particularly in data privacy and resource allocation. In the digital age, the usage of patient data for improving healthcare delivery and outcomes must navigate the stringent regulations set forth by data protection laws such as the General Data Protection Regulation (GDPR) and the Health Insurance Portability and Accountability Act (HIPAA). The imperative to safeguard patient privacy often clashes with the need to aggregate and analyze health data for research and predictive modeling purposes. Furthermore, NICUs face the ongoing challenge of efficiently allocating limited resources—such as specialized personnel, equipment, and beds—to meet the fluctuating care demands for premature neonates. These challenges underscore the necessity for innovative approaches to use data for predictive modeling without infringing on privacy regulations.

In response to these challenges, a compelling need arises for predictive models capable of operating within a decentralized, privacy-preserving data framework. Rhino Health’s FCP with edge computing and Federated Learning capabilities emerges as a promising solution, offering a paradigm wherein model training can occur across multiple sites without centralizing sensitive data. This approach aligns with the privacy-preserving mandates of contemporary data protection legislation and harnesses collective insights from diverse datasets across various NICUs. By synthesizing knowledge from decentralized data sources, Federated Learning models can provide robust, generalized predictions of LOS for premature neonates. Such models represent a significant advancement in neonatal care, promising to enhance resource utilization, care planning, and patient outcomes while protecting patient privacy.

Generative AI image

2. Rhino Health’s Federated Computing Platform Approach

Rhino Health’s FCP uses Federated Learning, an ML approach that trains algorithms across multiple decentralized servers while keeping data localized. This methodology is particularly beneficial in healthcare, where the protection of patient data is paramount. By using Federated Learning, healthcare institutions can collaborate on developing more accurate and generalized AI models without sharing sensitive patient data. This ensures compliance with stringent data privacy regulations and harnesses the collective intelligence of diverse data sets, leading to improved diagnostic tools and treatment plans.

The collaboration between the I-Medata AI Center at the Tel Aviv (Sourasky) Medical Center and Rhino Health was characterized by a shared commitment to exploring the frontiers of Federated Learning technology. Rhino Health provided the necessary Federated Computing infrastructure, and the development and refinement of the Quantile Regression model were carried out by the I-Medata AI Center’s teams. The I-Medata AI Center’s deep domain expertise was crucial to the success of this project. Together, these efforts resulted in a proof of concept to assess Rhino Health’s FCP capabilities.

The Federated Learning Quantile Regression model is an ML algorithm that runs on Rhino Health’s FCP to predict LOS for premature neonates in NICUs. The model was trained using a Federated Learning framework, using distributed data sources to train the algorithm without centralizing patient information. This Quantile Regression approach estimates the LOS distribution’s conditional median or other quantiles, offering a more nuanced understanding than average-based predictions. This method is particularly suited to healthcare data, which often exhibits skewed distributions, providing insights that can help optimize NICU resource allocation and improve neonatal care.

Rhino Health’s FCP uses the NVIDIA FLARE™ (NVIDIA Federated Learning Application Runtime Environment) framework and simplifies the transition of traditional ML workflows into the Federated Learning environment. The FCP reduces complexity by providing a unified interface, enabling developers to adapt and extend models such as Quantile Regression, Deep Learning, Tree-based models, etc., with greater ease and reduced risk of errors. Rhino Health’s engineering and data science teams worked with Ichilov’s data scientists and clinicians to efficiently federate and customize a Quantile Regression model, ensuring the use of privacy-enhancing techniques. This effort used Rhino Health FCP’s flexibility and extensibility to support various ML models tailored to meet customer-specific needs. Rhino Health’s FCP accelerated the model federation process through this sophisticated approach and overcame challenges in development, facilitating seamless adjustments to the codebase while maintaining high data privacy security standards and collaborative innovation.

3. Technological Significance and Innovation

Rhino Health’s approach includes simulating Federated Learning environments within single institutions. This approach serves as an evaluation mechanism for Federated Learning algorithms, offering institutions a means to assess Federated Learning capabilities in incorporating data from various sites in a privacy-preserving manner.

The Federated Learning Quantile Regression model runs in Rhino Health’s FCP and adheres to privacy considerations inherent to the Federated Learning approach, ensuring that data handling aligns with the principles like other Federated Learning models. By design, the model ensures that patient data remains within the institutional firewall, with only model parameters being exchanged. This approach keeps data local, thereby reducing the risk of exposure in a centralized repository and making it easier to comply with privacy regulations, offering an appealing solution for healthcare institutions focused on maintaining the confidentiality of sensitive patient information.

Rhino Health’s Federated Learning approach can transform healthcare AI by setting new privacy-preserving, collaborative model development standards. By facilitating secure, multi-institutional collaborations, Rhino Health’s FCP enables the development of AI models that are both highly accurate and broadly applicable across diverse patient populations. Furthermore, this approach democratizes access to cutting-edge healthcare innovations, allowing smaller institutions to contribute to and benefit from shared AI advancements. As such, Rhino Health is paving the way for a future where healthcare AI is more inclusive, effective, and secure.

4. Federated Learning Quantile Regression Model Outcomes

The study conducted by the I-Medata AI Center at the Tel Aviv (Sourasky) Medical Center and Rhino Health has demonstrated the feasibility and effectiveness of using a Federated Learning approach to predict the LOS for preterm newborns in NICU setting. The Quantile Regression model was trained on separate cohorts, each reflecting a unique method of data partitioning—centralized, random split, feature split, and label split.

The centralized approach, which used all available training data, set a benchmark with an R-squared (r2) value of 0.78705, a mean absolute error (MAE) of 6.092094, and a root mean square error (RMSE) of 10.993. The random split cohort closely mirrored these metrics, indicating that random data partitioning does not significantly degrade the model performance. In contrast, the feature split cohort, based on the median of the feature ‘weight at birth,’ and the label split cohort, based on the LOS label, exhibit diminished predictive capabilities as evidenced by lower r2 values and higher error metrics. Specifically, the feature split cohort resulted in an r2 of 0.693683, MAE of 7.017354, and RMSE of 13.1845, while the label split cohort further reduced r2 to 0.63486, increased MAE to 8.214184, and RMSE  to 14.39487.

The Federated Learning approach maintained performance parity with the local runs, thus validating the federated model’s ability to yield equivalent results to the traditional centralized training methods. This outcome is particularly noteworthy as it suggests that Federated Learning can achieve high-quality predictive modeling without compromising the privacy and security of sensitive patient data.

These results underscore the potential of Federated Learning in healthcare settings, where protecting patient data is paramount. They pave the way for further exploration into using federated models in various healthcare predictive tasks, with the promise of maintaining data sovereignty and privacy.

The findings from this experiment contribute to the growing body of evidence that Federated Learning is not only a viable option for sensitive environments but also an approach that can produce results on par with traditional centralized data training methods. This has profound implications for the future of AI in healthcare, where data privacy concerns are paramount, and the need for robust generalizable models is critical.

5. Conclusion

The successful collaboration between the I-Medata AI Center at the Tel Aviv (Sourasky) Medical Center and Rhino Health signifies a pivotal exploration into the capabilities of AI/ML and Federated Learning in healthcare. Implementing the Federated AI Quantile Regression Model served as another data point in validating the efficacy of Federated Learning en route to the broader goal of using technology for healthcare advancements. This project paves the way for future explorations into how healthcare data can be used for AI training in a secure, collaborative, and privacy-preserving manner.

Contact Rhino Health today to discover how our Federated Computing Platform offers a new era of collaborative innovation, ensuring data security and privacy-preserving compliance.