Zicheng (Jeremy) Xu

Zicheng (Jeremy) Xu

Works on LLMs & Medical AI

Publications

You can also find my articles on my Google Scholar profile.

Conference Papers

Self-ensemble: Mitigating Confidence Distortion for Large Language Models

Published in EMNLP Findings, 2025

Although Large Language Models (LLMs) perform well in general fields, they exhibit a confidence distortion problem on multi-choice question-answering (MCQA), particularly as the number of answer choices increases. Specifically, on MCQA with many choices, LLMs suffer from under-confidence in correct predictions and over-confidence in incorrect ones, leading to a substantially degraded performance. To solve this problem, we propose Self-ensemble in this work. Our method splits the choices into several groups and ensembles LLM predictions across these groups to reach a final decision. The advantage of Self-ensemble is its plug-and-play nature, where it can be integrated into existing LLM architecture based on a designed attention mask and positional encoding, without requiring labeled datasets for parameter tuning. Experimental results on three LLMs and datasets demonstrate that Self-ensemble comprehensively addresses the confidence distortion problem of LLMs, outperforming standard inference as well as baseline methods.

Download Paper

Journal Articles

Optimizing beat-wise input for arrhythmia detection using 1-D convolutional neural networks: A real-world ECG study

Published in Computer Methods and Programs in Biomedicine, 2025

Cardiac arrhythmias, characterized by irregular heartbeats, are difficult to diagnose in real-world scenarios. Machine learning has advanced arrhythmia detection; however, the optimal number of heartbeats for precise classification remains understudied. This study addresses this using machine learning while assessing the performance of arrhythmia detection across inter-patient and intra-patient conditions. Furthermore, the performance–resource trade-offs are evaluated for practical deployment in mobile health (mHealth) applications. Beat-wise segmentation and resampling techniques were utilized for preprocessing electrocardiography (ECG) signals to ensure consistent input lengths. A 1-D convolutional neural network was used to classify the eight multi-labeled arrhythmias. The dataset comprised real-world ECG recordings from the HiCardi wireless device alongside data from the MIT-BIH Arrhythmia database. Model performance was assessed through fivefold cross-validation under both inter-patient and intra-patient conditions.

Download Paper

Machine Learning Assisted Stroke Prediction in Mechanical Circulatory Support: Predictive Role of Systemic Mitochondrial Dysfunction

Published in ASAIO Journal, 2025

Stroke continues to be a major adverse event in advanced congestive heart failure (CHF) patients after continuous-flow left ventricular assist device (CF-LVAD) implantation. Abnormalities in mitochondrial oxidative phosphorylation (OxPhos) have been critically implicated in the pathogenesis of neurodegenerative diseases and cerebral ischemia. We hypothesize that prior stroke may be associated with systemic mitochondrial OxPhos abnormalities, and impaired more in post-CF-LVAD patients with risk of developing new stroke. We studied 50 CF-LVAD patients (25 with prior stroke, 25 without); OxPhos complex proteins (complex I [C.I]–complex V [C.V]) were measured in blood leukocytes. Both at baseline (pre-CF-LVAD) and postoperatively (post-CF-LVAD), the prior-stroke group had significantly lower C.I, complex II (C.II), complex IV (C.IV), and C.V proteins when compared to the no-prior-stroke group. Oxidative phosphorylation proteins were significantly decreased in prior-stroke group at post-CF-LVAD compared to pre-CF-LVAD. Machine learning Least Absolute Shrinkage and Selection Operator (LASSO) and Random Forest modeling identified six prognostic factors that predicted postoperative stroke with an area under the receiver operating characteristic (ROC) curve (AUC) of 0.93. Oxidative phosphorylation protein reduction appeared to be associated with the new stroke after implantation. Our study found for the first time the existence of mitochondrial dysfunction at the peripheral level in CHF patients with prior ischemic stroke even before CF-LVAD implantation. The changes in OxPhos protein expression could serve as biomarkers in predicting new post-CF-LVAD strokes.

Download Paper

Preoperative brain volume loss is associated with postoperative delirium in advanced heart failure patients supported by left ventricular assist device

Published in Nature Scientific Reports, 2025

Delirium is a common neurological complication in patients with advanced heart failure (ADHF) following left ventricular assist device (LVAD) implantation, significantly impacting recovery. This study aimed to analyze non-contrast computed tomography (CT) scans of the brain in ADHF patients undergoing LVAD implantation to determine the association between pre-existing brain atrophy and postoperative delirium. A study involving 166 ADHF patients was conducted from March 2020 to July 2023. Non-contrast CT scans were analyzed using advanced quantitative neuroimaging techniques before implantation. The primary marker assessed was the lateral ventricle fraction (LVF), with secondary markers including cortical gray matter fraction (cGMF), white matter fraction (WMF), basal ganglia fraction (BGF), and thalamus fraction (TLF). A total of 56 patients (33%) experienced postoperative delirium within two weeks of implantation. Patients with delirium were older and exhibited greater brain atrophy, indicated by higher LVF and lower cGMF, WMF, BGF, and TLF values. The occurrence of delirium was strongly associated with age, and ventricular enlargement, primarily in the lateral ventricles. LVF effectively predicted delirium development, regardless of age. Preoperative brain volumetric analysis, particularly of the lateral ventricles, may be crucial in identifying patients at risk for postoperative delirium, enhancing postoperative management, and improving outcomes for LVAD recipients.

Download Paper

Hierarchical deep learning for autonomous multi-label arrhythmia detection and classification on real-world wearable electrocardiogram data

Published in Digital Health, 2024

Arrhythmia detection and classification are challenging because of the imbalanced ratio of normal heartbeats to arrhythmia heartbeats and the complicated combinations of arrhythmia types. Arrhythmia classification on wearable electrocardiogram monitoring devices poses a further unique challenge: unlike clinically used electrocardiogram monitoring devices, the environments in which wearable devices are deployed are drastically different from the carefully controlled clinical environment, leading to significantly more noise, thus making arrhythmia classification more difficult. We propose a novel hierarchical model based on CNN+BiLSTM with Attention to arrhythmia detection, consisting of a binary classification module between normal and arrhythmia heartbeats and a multi-label classification module for classifying arrhythmia events across combinations of beat and rhythm arrhythmia types. We evaluate our method on our proprietary dataset and compare it with various baselines, including CNN+BiGRU with Attention, ConViT, EfficientNet, and ResNet, as well as previous state-of-the-art frameworks. Our model outperforms existing baselines on the proprietary dataset, resulting in an average accuracy, F1-score, and AUC score of 95%, 0.838, 0.906 for binary classification, and 88%, 0.736, 0.875 for multi-label classification.

Download Paper