<p>Neurological disorders globally affect millions of people, a major burden, both personally and financially for individuals, as well as for society. Although deep learning models provide accurate diagnoses in laboratory settings with curated data. The models do not perform when they are used on newer portable, user friendly, Portable and clinical EEG devices can produce varying amounts of channels of EEG recordings to produce a greater amount of real-world EEG recordings. In addition to there being fewer channels than with traditional EEG devices this reliability issue prevents the clinician from being able to use the output of the deep learning model to make an informed decision, because there is no way to provide the clinician with the uncertainty associated with the output using currently available standard modeling. Therefore, this work’s objectives are to present a framework called a reliability-based LSTM, a multi-task learning system of analysing the EEG; for use in a clinical setting. the analysis has been modified to using a dual-head LSTM with three qualitatively different options for estimating the outputs easy measure High Reliability Score (HRS), temperature scaling, and conformal prediction. This integration will allow a model to continue to provide a high degree of diagnostic confidence and produce appropriately calibrated confidence estimates for risk-based decision making. The proposed model achieved a seizure detection accuracy rate of 96.66%, while also providing appropriately calibrated results as evidenced by near perfect performance on reliability diagrams and the Risk-Coverage curve. The HRS allowed the authors to risk stratify using a mean HRS of 0.78 and a high confidence (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\ge 0.8$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>≥</mo> <mn>0.8</mn> </mrow> </math></EquationSource> </InlineEquation>) rate covering 72.88%. This highlighted the pathway available for partially automating clinical surveillance application where a high-confidence prediction could initiate an immediate alert and cases of low confidence would be flagged for expert review. In conclusion, this project provided precise diagnoses of EEG data in combination with reliable, actionable measures of reliability from which, high performance models can be transformed into clinically reliable tools suitable for real world use. Our experiments were conducted using CHB-MIT as the clinical data set. The reliability framework that has been developed can potentially be translated to noisier settings with fewer channels of EEG; however, this has yet to be verified in future work with real mobile EEG devices using the proposed framework.</p>

Reliability-quantified multi-task learning: Dual-head architectures of risk-awareness for EEG devices

期刊论文
发表日期：

2026年7月13日

查看全文

Wenjuan Wu,
Mohsin Hasan,
Xufeng Zhao

摘要

Neurological disorders globally affect millions of people, a major burden, both personally and financially for individuals, as well as for society. Although deep learning models provide accurate diagnoses in laboratory settings with curated data. The models do not perform when they are used on newer portable, user friendly, Portable and clinical EEG devices can produce varying amounts of channels of EEG recordings to produce a greater amount of real-world EEG recordings. In addition to there being fewer channels than with traditional EEG devices this reliability issue prevents the clinician from being able to use the output of the deep learning model to make an informed decision, because there is no way to provide the clinician with the uncertainty associated with the output using currently available standard modeling. Therefore, this work’s objectives are to present a framework called a reliability-based LSTM, a multi-task learning system of analysing the EEG; for use in a clinical setting. the analysis has been modified to using a dual-head LSTM with three qualitatively different options for estimating the outputs easy measure High Reliability Score (HRS), temperature scaling, and conformal prediction. This integration will allow a model to continue to provide a high degree of diagnostic confidence and produce appropriately calibrated confidence estimates for risk-based decision making. The proposed model achieved a seizure detection accuracy rate of 96.66%, while also providing appropriately calibrated results as evidenced by near perfect performance on reliability diagrams and the Risk-Coverage curve. The HRS allowed the authors to risk stratify using a mean HRS of 0.78 and a high confidence ( $\ge 0.8$ $\geq 0.8$ ) rate covering 72.88%. This highlighted the pathway available for partially automating clinical surveillance application where a high-confidence prediction could initiate an immediate alert and cases of low confidence would be flagged for expert review. In conclusion, this project provided precise diagnoses of EEG data in combination with reliable, actionable measures of reliability from which, high performance models can be transformed into clinically reliable tools suitable for real world use. Our experiments were conducted using CHB-MIT as the clinical data set. The reliability framework that has been developed can potentially be translated to noisier settings with fewer channels of EEG; however, this has yet to be verified in future work with real mobile EEG devices using the proposed framework.