Luo He

Luo He

Kuaishou (China)
4 previous affiliations
3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Luo He focuses on healthcare analytics, medical image diagnosis, and predictive modeling for clinical outcomes. Early work explored the socioeconomic factors influencing hospital readmissions and applied deep learning to ophthalmic disease detection. More recently, the focus has shifted toward developing attention-based recurrent neural networks that leverage longitudinal physical examination data to predict chronic conditions.

Research Trajectory

The trajectory of Luo He’s research is characterized by the application of data mining and deep learning techniques to diverse forms of medical data, ranging from structured hospital records to unstructured medical images. The earliest available work in the corpus, emerging around 2017, demonstrates a dual interest in health economics and computer vision applied to clinical diagnostics.

In 2017, He led research investigating the socioeconomic and demographic drivers of hospital readmission rates, a critical metric for both care quality and healthcare expenditure. As the first author of How Out-of-Pocket Ratio Influences Readmission: An Analysis Based on Front Sheet of Inpatient Medical Record, He and colleagues analyzed a substantial dataset comprising inpatient records from 150 hospitals in Beijing collected between 2012 and 2014. The study specifically examined the relationship between a patient's out-of-pocket financial burden and their likelihood of readmission. Because inpatient records contain a vast array of distinct diseases and surgical operations, the research utilized data mining methodologies to isolate discriminative features, focusing particularly on records with primary diagnoses related to rectal conditions. This work highlighted the intersection of financial policy, patient demographics, and clinical outcomes.

Concurrently in 2017, He participated as the second author in a collaborative effort to improve automated medical image analysis. The paper Deep Learning Through Two-Branch Convolutional Neuron Network for Glaucoma Diagnosis addressed a fundamental challenge in medical machine learning: the scarcity of large, annotated datasets required to train deep neural networks effectively. To overcome the limitations of small datasets and constrained computing resources, the research team proposed a novel two-branch convolutional neural network (CNN) architecture designed to detect glaucoma from retinal images. By explicitly incorporating domain knowledge into the network's design, the model was able to achieve robust diagnostic performance without relying on the massive data volumes typically required by standard computer vision models.

Following these initial publications, the evidence shows a progression toward more complex, longitudinal predictive modeling. In 2022, He returned as first author to publish A Multi-attention Collaborative Deep Learning Approach for Blood Pressure Prediction. This research tackled the challenge of forecasting chronic conditions using multi-year physical examination and laboratory results. A primary technical hurdle in this domain is the sparsity of data, as individual patients often have short or irregular examination histories. To address this, He developed the Multi-attention Collaborative Deep Learning model (MAC-LSTM). Built upon a Long Short-Term Memory network foundation, the architecture integrated three distinct types of attention mechanisms to improve both the accuracy and the clinical explainability of the predictions. Furthermore, the model incorporated a collaborative filtering approach, leveraging health trajectories from similar users to augment the predictive power for patients with limited historical data. This approach demonstrated significant reductions in predictive error compared to existing baseline models.

Current Directions

Based on the most recent paper in the evidence from 2022, Luo He’s current research directions center on advanced deep learning architectures for longitudinal health data. The primary focus is on overcoming the inherent challenges of real-world clinical datasets, such as sparse observation points and short patient histories.

Recent work investigates the integration of collaborative filtering techniques with recurrent neural networks (such as LSTMs) to pool predictive signals across similar patient cohorts. Additionally, there is a strong emphasis on model interpretability. By embedding multi-attention mechanisms into predictive models, the research aims to ensure that complex algorithms not only forecast clinical metrics—such as blood pressure—with high accuracy, but also provide explainable outputs that can be trusted and understood by medical practitioners.

Profile updated

See something wrong? Report a mistake