Machine Learning, and in particular Federated Machine Learning, opens new perspectives in terms of medical research and patient care. Although Federated Machine Learning improves over centralized Machine Learning in terms of privacy, it does not provide provable privacy guarantees. Furthermore, Federated Machine Learning is quite expensive in term of bandwidth consumption as it requires participant nodes to regularly exchange large updates. This paper proposes a bandwidth-efficient privacy-preserving Federated Learning that provides theoretical privacy guarantees based on Differential Privacy. We experimentally evaluate our proposal for in-hospital mortality prediction using a real dataset, containing Electronic Health Records of about one million patients. Our results suggest that strong and provable patient-level privacy can be enforced at the expense of only a moderate loss of prediction accuracy.

https://dl.acm.org/doi/pdf/10.1145/3450439.3451859

Privacy-preserving and bandwidth-efficient federated learning: an application to in-hospital mortality prediction

Background: Early warning tools identify patients at risk of deterioration in hospitals. Electronic medical records in hospitals offer real-time data and the opportunity to automate early warning tools and provide real-time, dynamic risk estimates.
Objective: This review describes published studies on the development, validation, and implementation of tools for predicting patient deterioration in general wards in hospitals.
Methods: An electronic database search of peer reviewed journal papers from 2008-2020 identified studies reporting the use of tools and algorithms for predicting patient deterioration, defined by unplanned transfer to the intensive care unit, cardiac arrest, or death. Studies conducted solely in intensive care units, emergency departments, or single diagnosis patient groups were excluded.
Results: A total of 46 publications were eligible for inclusion. These publications were heterogeneous in design, setting, and outcome measures. Most studies were retrospective studies using cohort data to develop, validate, or statistically evaluate prediction tools. The tools consisted of early warning, screening, or scoring systems based on physiologic data, as well as more complex algorithms developed to better represent real-time data, deal with complexities of longitudinal data, and warn of deterioration risk earlier. Only a few studies detailed the results of the implementation of deterioration warning tools.
Conclusions: Despite relative progress in the development of algorithms to predict patient deterioration, the literature has not shown that the deployment or implementation of such algorithms is reproducibly associated with improvements in patient outcomes. Further work is needed to realize the potential of automated predictions and update dynamic risk estimates as part of an operational early warning system for inpatient deterioration.

Predicting Patient Deterioration: A Review of Tools in the Digital Hospital Setting

Federated Learning allows distributed entities to train a common model collaboratively without sharing their own data. Although it prevents data collection and aggregation by exchanging only parameter updates, it remains vulnerable to various inference and reconstruction attacks where a malicious entity can learn private information about the participants' training data from the captured gradients. Differential Privacy is used to obtain theoretically sound privacy guarantees against such inference attacks by noising the exchanged update vectors. However, the added noise is proportional to the model size which can be very large with modern neural networks. This can result in poor model quality. In this paper, compressive sensing is used to reduce the model size and hence increase model quality without sacrificing privacy. We show experimentally, using 2 datasets, that our privacy-preserving proposal can reduce the communication costs by up to 95% with only a negligible performance penalty compared to traditional non-private federated learning schemes.

Compression Boosts Differentially Private Federated Learning

Federated Learning allows distributed entities to train a common model collaboratively without sharing their own data. Although it prevents data collection and aggregation by exchanging only parameter updates, it remains vulnerable to various inference and reconstruction attacks where a malicious entity can learn private information about the participants' training data from the captured gradients. Differential Privacy is used to obtain theoretically sound privacy guarantees against such inference attacks by noising the exchanged update vectors. However, the added noise is proportional to the model size which can be very large with modern neural networks. This can result in poor model quality. In this paper, compressive sensing is used to reduce the model size and hence increase model quality without sacrificing privacy. We show experimentally, using 2 datasets, that our privacy-preserving proposal can reduce the communication costs by up to 95 % with only a negligible performance penalty compared to traditional non-private federated learning schemes.

https://hal.archives-ouvertes.fr/hal-03066941/document

The abundant availability of health-care data calls for effective analysis methods to help medical experts gain a better understanding of their patients and their health. The focus of existing work has been largely on prediction. In this paper, we introduce Core, a framework for cohort “representation” and “exploration.” Our contributions are twofold: First, we formalize cohort representation as the problem of aggregating the trajectories of its patients. This problem is challenging because cohorts often consist of hundreds of patients who underwent medical actions of various types at different points in time. We prove that producing a representative cohort trajectory is NP-complete with a reduction in the multiple sequence alignment problem. We propose a heuristic that extends the Needleman–Wunsch algorithm for sequence matching to handle temporal sequences. To further improve cohort representation efficiency, we introduce “trajectory families” and “stratified sampling.” Our second contribution is formalizing the problem of cohort exploration as finding a set of cohorts that are similar to a cohort of interest and that maximize entropy. This problem is challenging because the potential number of similar cohorts is huge. We prove NP-completeness with a reduction in the maximum edge subgraph problem. To address complexity, we develop a multi-staged approach based on limiting the search space to “contrast cohorts.” To speed up the computation of cohort similarity, we use “event sets” that are inspired from the double dictionary encoding proposed for keyword search. Moreover, we explore the usefulness and efficiency of Core using an extensive set of qualitative and quantitative experiments on two real health-care datasets. In a user study with medical experts, we show that Core reduces time-to-insight from hours to seconds and helps them find better insights than baseline approaches. Moreover, we show that the obtained cohort representations offer the right trade-off between quality and performance. We study the benefits of trajectory families and stratified sampling for cohort representation and show their applicability on large and heterogeneous cohorts. We also show the benefit of event sets for cohort exploration in providing interactive performance.

Cohort analytics: efficiency and applicability

Early identification of patients at risk of developing complications during their hospital stay is currently one of the most challenging issues in healthcare. Complications include hospital-acquired infections, admissions to intensive care units, and in-hospital mortality. Being able to accurately predict the patients' outcomes is a crucial prerequisite for tailoring the care that certain patients receive, if it is believed that they will do poorly without additional intervention. We consider the problem of complication risk prediction, such as inpatient mortality, from the electronic health records of the patients. We study the question of making predictions on the first day at the hospital, and of making updated mortality predictions day after day during the patient's stay. We develop distributed models that are scalable and interpretable. Key insights include analysing diagnoses known at admission and drugs served, which evolve during the hospital stay. We leverage a distributed architecture to learn interpretable models from training datasets of gigantic size. We test our analyses with more than one million of patients from hundreds of hospitals, and report on the lessons learned from these experiments.

/pdf/scalable-and-interpretable-predictive-models-for-electronic-oz57i0hzb7.pdf

Scalable and Interpretable Predictive Models for Electronic Health Records

We consider the optimization of recursive queries over graphs, and more specifically t he p roblem o f e fficiently enumerating recursive query plans. For a given query, the enumeration of logical plans – and in particular join enumeration – is an ex-tensively researched topic. However, query operators such as union and recursion have been largely neglected in enumeration techniques. We propose an extension of the logical query DAG (LQDAG) used in modern query optimizers with the support of recursion, and develop appropriate techniques to enumerate recursive terms efficiently. Ou r LQ DAG extension is based on the fixpoint o perator r ecently i ntroduced i n the 𝜇 -relational algebra. We show how to adapt transformation rules for individual fixpoint terms so that they can apply on a group of terms at once. The main benefit i s t o e nable a much more efficient ex ploration of re cursive qu ery pl an sp aces. We report on practical experiments with graph queries over real and synthetic graphs of varying size. Experimental results il-lustrate the benefits of exploring faster larger query execution plan spaces.

Optimizing Enumeration of Recursive Plans in Transformation-based Query Optimizers

We demonstrate a system for recursive query answering over property graphs. The novelty of the system resides in its ability to optimize and efficiently answer recursive path patterns in queries for property graphs. The system is based on a complete implementation of the recent µ -recursive relational algebra. It also includes parsers and compilers adapted for property graphs so that one can formulate, optimize and answer queries that navigate recursively along paths in property graphs. We demonstrate the system on three real datasets, including the exploration of chains of drug interactions.

Amela Fejza

Papers

Scalable and Interpretable Predictive Models for Electronic Health Records

Optimizing Enumeration of Recursive Plans in Transformation-based Query Optimizers

Exploring Labeled Graphs with Recursive Path Patterns