Scholarlink

<p>In this paper, we investigate two problems faced in detecting/predicting intracranial hypertension (ICH) onsets. The first challenge is that intracranial pressure (ICP) can be measured only invasively, drilling a hole in patients skull. To tackle this challenge, we propose a novel metadata supported scale space attention (SSA) network to accurately and <i>non-invasively</i> predict intracranial hypertension (ICH) onsets, relying on supporting multi-variate data streams, including electroencephalogram (EEG), arterial blood pressure (ABP) and electrocardiogram (ECG). In particular, we propose a coupled network, integrating a convolutional neural network with <i>scale space attention</i> (CNN-SSA) and a long short-term memory (LSTM) network. This, however, opens up a new challenge—EEG readings at different health centers may rely on hardware with different numbers of sensors, which would render models not shareable across centers. To tackle this second challenge, we propose two alternative techniques: the first alternative relies on a multi-variate multi-scale neural network (M2NN) to impute missing EEG channels; the second alternative, on the other hand, identifies a core subset of EEG channels, shared by different centers, that can still provide accurate ICH predictions. We conduct rigorous experiments to substantiate our proposed method’s effectiveness—the results show that ICH can be predicted accurataly by relying on supporting data modalities, if we can properly attend the data at multiple scales and across modalities. Additionally, an ablation study demonstrates that the various components of our approach are essential in providing accurate ICH predictions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Metadata supported scale space attention networks for multivariate timeseries prediction

  • 期刊论文

  • 发表日期：

    2025年8月19日

查看全文
  • Manjusha Ravindranath,
  • K. Selçuk Candan,
  • Brian Appavu

摘要

In this paper, we investigate two problems faced in detecting/predicting intracranial hypertension (ICH) onsets. The first challenge is that intracranial pressure (ICP) can be measured only invasively, drilling a hole in patients skull. To tackle this challenge, we propose a novel metadata supported scale space attention (SSA) network to accurately and non-invasively predict intracranial hypertension (ICH) onsets, relying on supporting multi-variate data streams, including electroencephalogram (EEG), arterial blood pressure (ABP) and electrocardiogram (ECG). In particular, we propose a coupled network, integrating a convolutional neural network with scale space attention (CNN-SSA) and a long short-term memory (LSTM) network. This, however, opens up a new challenge—EEG readings at different health centers may rely on hardware with different numbers of sensors, which would render models not shareable across centers. To tackle this second challenge, we propose two alternative techniques: the first alternative relies on a multi-variate multi-scale neural network (M2NN) to impute missing EEG channels; the second alternative, on the other hand, identifies a core subset of EEG channels, shared by different centers, that can still provide accurate ICH predictions. We conduct rigorous experiments to substantiate our proposed method’s effectiveness—the results show that ICH can be predicted accurataly by relying on supporting data modalities, if we can properly attend the data at multiple scales and across modalities. Additionally, an ablation study demonstrates that the various components of our approach are essential in providing accurate ICH predictions.