Global AI Impact Commons
AI-Enabled Delirium Detection Using Portable EEG and Topological Data Analysis
Japan
Healthcare
Implementing Organisation
Ministry of Foreign Affairs
Japan, Tokyo, 2-2-1 Kasumigaseki Chiyoda-ku
Implementing Point of Contact
Hiroshi Kawamura
Deputy Director, Economic Diplomacy Strategy Division
Contributor of the Impact Story
Japan
Year of implementation
2016
Problem statement
Delirium is one of the most common, dangerous, and underdiagnosed conditions affecting hospitalized older adults worldwide. In routine clinical practice, detection relies largely on repeated questionnaire-based assessments, which are difficult to apply consistently by busy healthcare staff and show poor sensitivity in real-world settings (reported 38–47% in busy clinical environments, including intensive care units). As a result, delirium is frequently missed or detected late, despite being strongly associated with mortality, prolonged hospitalization, institutionalization after discharge, and long-term cognitive decline. The burden extends beyond individual patients, imposing substantial systemic costs, reported to exceed USD 150 billion annually in the United States alone. While conventional multi-lead EEG can detect delirium-related diffuse brain slowing, it is impractical for large-scale screening due to equipment complexity, specialist staffing requirements, and delays in interpretation. Earlier EEG-based algorithmic approaches, including simplified spectral methods, demonstrated feasibility but were limited in accuracy due to reliance on frequency features alone. This case addresses this global gap by introducing Topological Data Analysis (TDA), a fundamentally different and interpretable AI approach that captures global temporal irregularities in EEG signals, combined with a simplified portable EEG method, to enable objective, scalable, and clinically practical delirium detection using real-world hospital data.
Impact story details
The relevant division is the Economic Diplomacy Strategy Division, Ministry of Foreign Affairs. Their stated mission is to realise economic growth and social good through economic diplomacy strategy.
AI Technology Used
Topological Data Analysis (persistent homology) applied to portable EEG time-series data, extending prior spectral EEG algorithms by capturing global temporal and structural signal.
Key Outcomes
Accuracy
Quality Improvement, Efficiency
Productivity, Resilience
Risk Reduction
Narrative Outcome
This technology uses portable EEG and topological data analysis to detect delirium in hospitalised older adults. Traditional questionnaire-based screening misses the majority of cases in busy clinical environments. The AI-enhanced approach significantly improves diagnostic accuracy, with detection rates improving substantially across independent cohorts and devices. Improved accuracy also means fewer false positives. The system enables objective, consistent screening that can scale without relying on specialist-heavy workflows
Impact Metrics
Improved Area Under the ROC Curve (AUC) with TDA v. the traditional BSEEG was documented across independent cohorts and devices. AUC measures how well the technology distinguishes between patients with and without delirium, and hence, an improved AUC signifies greater accuracy in correct identification of delirium cases.
Baseline Value
While using BSEEG, Cohort 1 (Fp1-AI electrode placement) recorded an AUC of 0.72, Cohort 1 (Fp2-A2 electrode placement) recorded an AUC of 0.76, and Cohort 2 (Fp1-A1 electrode placement) recorded an AUC of 0.67. Area Under the ROC Curve (AOC)
Post-Implementation
While using TDA, Cohort 1 (Fp1-AI electrode placement) recorded an improved AUC of 0.82, Cohort 1 (Fp2-A2 electrode placement) recorded an improved AUC of 0.84, and Cohort 2 recorded an improved AUC of 0.78. Area Under the ROC Curve (AOC)
Statistically significant and clincially meaningful improvements in delirium detection specificity were also recorded using the DeLongtest at a fixed sensitivity i.e. a fixed accuracy rate.
Baseline Value
0
Post-Implementation
At a fixed sensitivity of 0.80 (80%), the use of TDA was seen to improve specificity in delirium detection by more than 16% to 18%, materially reducing false positivites. This demostrated identification of delirium cases that had been missed by spectral features alone.
Percentage of EEG (electroencephalogram) recordings retained after preprocessing for TDA analysis was also recorded as a primary clinical performance metric
Baseline Value
0
Post-Implementation
0
The time spent per screening session at bedside of patients was also recorded as a primary clinical performance metric.
Baseline Value
0
Post-Implementation
0
The repeatability rate of TDA at which clinicians could obtain a clean or usable EEG signal quality for analysis was also recorded as a primary clinical performance metric.
Baseline Value
0
Post-Implementation
0
Implementation Context
University of Iowa Hospitals and Clinics in the United States
The target population is a sample size of 480 patients primarily hospitalized adults aged 55 years or older in clinical setting such as the medicine floor, orthopedics floor, and emergency departments. The target demographic is predominantly non-Hispanic White per self-reporting; including both male and female participants.
Key Partnerships
Fujitsu Laboratories who provided the Topological Data Analysis (TDA) technology and algorithmic expertise, Delight Health, which is the startup collaborator applying TDA to delirium detection using EEG founded by a clinical leader, University of Iowa hospitals and clinics, which are the clinical sites for patient recruitment, data collection, and validation, Board-certified psychiatrists provided clinical overight and reviewed cases for final delirium case definition, with raters blinded to EEG scoring
Replicability & Adaptation
1. The approach is designed around minimal hardware requirements and objective signal-based analysis, making it broadly applicable across healthcare systems 2. Replication requires contextual adaptation, including validation across diverse populations, parameter tuning for different EEG devices, and computational optimization 3. These requirements are explicitly documented in the study and are typical for clinically deployed AI 4. Validate across more diverse populations and institutions to improve generalizability 5. Reduce sample loss caused by strict signal quality criteria through improved preprocessing balance 6. Standardize or auto-tune TDA parameters across devices and sampling rates 7. Optimize computational speed for near-real-time bedside scoring 8. Continue mitigation of EMG contamination overlapping EEG frequency bands
1. Portable single- or dual-channel EEG devices and consumables 2. Signal preprocessing and artifact mitigation pipeline 3. TDA computation capability (currently PC-based; target is edge/handheld) 4. Clinical assessment workflow for ground truth labeling (CAM-ICU, DRS, DOSS, MoCA, documentation review) 5. Data governance approvals (IRB and informed consent process used in the study) 6. Training for non-expert staff to place electrodes and operate the device
Supporting Materials
* The data presented is self-reported by the respective organisations. Readers should consult the original sources for further details.