All Projects → zhengqi98 → Hefei_ECG_TOP1

zhengqi98 / Hefei_ECG_TOP1

Licence: other
“合肥高新杯”心电人机智能大赛 —— 心电异常事件预测 TOP1 Solution

Programming Languages

python
139335 projects - #7 most used programming language

Projects that are alternatives of or similar to Hefei ECG TOP1

ECG-acquisition-classification
Single Lead ECG signal Acquisition and Arrhythmia Classification using Deep Learning
Stars: ✭ 41 (-62.39%)
Mutual labels:  ecg, ecg-classification
automatic-ecg-diagnosis
Scripts and modules for training and testing neural network for ECG automatic classification. Companion code to the paper "Automatic diagnosis of the 12-lead ECG using a deep neural network".
Stars: ✭ 156 (+43.12%)
Mutual labels:  ecg, ecg-classification
The-Purchase-and-Redemption-Forecast-Challenge-baseline
天池“资金流入流出预测——挑战baseline”的解决方案,线上效果143.5
Stars: ✭ 78 (-28.44%)
Mutual labels:  competition, data-mining
ECG analysis
No description or website provided.
Stars: ✭ 32 (-70.64%)
Mutual labels:  data-mining, ecg
tool-competition-av
The repository hosts the code for the SBST CPS Tool competition for testing autonomous cars.
Stars: ✭ 16 (-85.32%)
Mutual labels:  competition
HFT-Prediction
Machine learning approach to high frequency trading, MLP & RNN used
Stars: ✭ 19 (-82.57%)
Mutual labels:  data-mining
PhysioNet
Deep learning based ECG classification
Stars: ✭ 15 (-86.24%)
Mutual labels:  ecg
2018-JData-Unicom-RiskUser
2018-JData-联通-基于移动网络通讯行为的风险用户识别:Baseline 0.77
Stars: ✭ 20 (-81.65%)
Mutual labels:  data-mining
fhq-server
This is an open source platform for competitions of computer security.
Stars: ✭ 33 (-69.72%)
Mutual labels:  competition
ECG-Heartbeat-Classification-seq2seq-model
Inter- and intra- patient ECG heartbeat classification for arrhythmia detection: a sequence to sequence deep learning approach
Stars: ✭ 125 (+14.68%)
Mutual labels:  ecg
candis
🎀 A data mining suite for gene expression data.
Stars: ✭ 28 (-74.31%)
Mutual labels:  data-mining
FEATHER
The reference implementation of FEATHER from the CIKM '20 paper "Characteristic Functions on Graphs: Birds of a Feather, from Statistical Descriptors to Parametric Models".
Stars: ✭ 34 (-68.81%)
Mutual labels:  data-mining
emperor-os
(new released v2.5 LTS.2022-06-25) It has focused on developing an All in One operating system for programming, designing and data science.Emperor-OS has over 500 apps and important tools
Stars: ✭ 32 (-70.64%)
Mutual labels:  data-mining
tree-hugger
A light-weight, extendable, high level, universal code parser built on top of tree-sitter
Stars: ✭ 96 (-11.93%)
Mutual labels:  data-mining
ifunny
Решение для конкурса FunCode Java/Kotlin Challenge
Stars: ✭ 25 (-77.06%)
Mutual labels:  competition
chainRec
Mengting Wan, Julian McAuley, "Item Recommendation on Monotonic Behavior Chains", in Proc. of 2018 ACM Conference on Recommender Systems (RecSys'18), Vancouver, Canada, Oct. 2018.
Stars: ✭ 52 (-52.29%)
Mutual labels:  data-mining
travelling-salesman
Rules for Kiwi.com travelling salesman competition
Stars: ✭ 14 (-87.16%)
Mutual labels:  competition
XCloud
Official Code for Paper <XCloud: Design and Implementation of AI Cloud Platform with RESTful API Service> (arXiv1912.10344)
Stars: ✭ 58 (-46.79%)
Mutual labels:  data-mining
Loan-Approval-Prediction
Loan Application Data Analysis
Stars: ✭ 61 (-44.04%)
Mutual labels:  data-mining
B2P-Penicilin-Tablut-AI
Java-based artificial intelligence that plays Tablut using Ashton's rules
Stars: ✭ 27 (-75.23%)
Mutual labels:  competition

“合肥高新杯”心电人机智能大赛——心电异常事件预测冠军解决方案

赛题地址

https://tianchi.aliyun.com/competition/entrance/231754/introduction

开源baseline

感谢比赛期间JavisPeng开源的优质baseline

比赛分数

赛题回顾

  • 问题描述:依据心电图机8导联的数据和年龄、性别特征,预测心电异常事件
  • 比赛数据:32142条初赛数据和20036条复赛数据(初赛数据有重复,初复赛标签分布差异大)
  • 评价指标:MicroF1

数据分析

  • 标签相关性:计算公式为两标签交集数量除以两标签并集数量,0表示完全互斥。该相关性也可视为一种“特征”。

  • 不同导联节拍一致:尖峰位置一致。

  • 不同导联十分相似:将不同导联画在同一坐标轴上,可看出相似性。如何构建模型以利用这种相似性最为关键的思路。

模型构建

针对多导联心电图分类任务,我创新地提出一种网络结构,将其称之为ECGNet: Multi-scale ResNet for Multi-lead ECG Data。该模型是本次比赛的致胜关键。(细节可见PPT)

模型融合

模型融合阶段效果提升,我认为主要有两点原因:

  1. 充分利用初赛和复赛的数据
  2. “隐含”地利用了不同标签的相关性

不足之处:

  1. 模型缺乏多样性
  2. 没有用到传统特征和树模型

此外,植物提出的嫁接学习也是种很有意思的思路。

效果评估

线下对20类标签用单模型评估效果,反正比我自己判断的(仅限电轴偏转方向)好得多。

经验总结

  • 充分利用提供的数据很重要,尤其分布差异很大时。
  • 多去思考多去尝试。

在处理多导联心电图数据时:

  • 对不同导联应用相同的卷积核,能在减小参数量的同时,很好地提升模型的效果。
  • 采用多尺度网络能捕捉不同尺度的特征,较好地提升模型的效果。
  • 网络初期可以采用较大的卷积核,后期可以采用较小的卷积核。
  • BN-ReLU-Conv要优于Conv-BN-ReLU。
  • Squeeze-and-excitation结构也能提升模型的效果。
  • 传统特征也很重要。

失败尝试

以下是一些失败的尝试(不代表这些方法真的不行,也许是我的打开方式有问题):

  • 傅里叶变换
  • 小波变换
  • 频谱图
  • DenseNet
  • EfficientNet
  • Attention
  • LSTM
  • 滑动窗口
  • 去噪
  • 特征工程

总之还是要多思考多去尝试吧,没有什么事能一帆风顺的。

赛后感想

本次比赛收获颇丰,除了实质性奖励以外,还锻炼了我赛题思考、数据分析、模型构建、论文阅读、编程实现以及答辩的能力,且与其他选手交流了一些有趣的思路。
客观上来看,本次比赛有很大的运气成分的,自己还有很多不足。总之,继续努力,再接再厉吧。
单人参赛好累啊,还有复现阶段需要抓紧时间,我差点没在期限内整出来。

PS

很多细节我没有详细介绍,有兴趣地可以看PPT或答辩视频。

答辩视频

https://tianchi.aliyun.com/course/video?liveId=41127

Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].