Johns Hopkins LSTM BCI lets a man with ALS spell for 90 days

AI system

System: Click-based electrocorticographic brain-computer interface with an LSTM click detector (NCT03567213)
Model: A long short-term memory network with 25 hidden units and two fully connected layers.
Release: Exact release not established.
Application developer: Johns Hopkins University (Crone lab and co-authors)
Role: The model detected intended clicks in cortical recordings.
Configuration: Researchers implemented the network with Keras and TensorFlow 2.8.0.
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: High-density ECoG over sensorimotor cortex. LSTM click detector trained on under 44 minutes of data. Switch-scanning speller interface.
Human oversight: Not assessed
Relation to AI: The LSTM click detector is the decoding element. Its stability without retraining is the paper's main result. The source is a peer-reviewed report by the developers, with no outside corroboration and no use outside the laboratory.
AI contribution: The LSTM click detector is the decoding element. Its stability without retraining is the paper's main result. The source is a peer-reviewed report by the developers, with no outside corroboration and no use outside the laboratory.

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: median spelling rate of 10.2 characters per minute
Comparison: the Utrecht fully implanted switch-scan BCI reported by Vansteensel et al. (four contacts over hand motor cortex)
Results: Reference: a spelling rate of 1.8 letters per minute with a latency of at least 1 s per click. Result: a spelling rate of 10.2 CCPM and 0.68 s latency.
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Participants: Developer: Johns Hopkins University (Crone lab and co-authors). Operator: Johns Hopkins University (laboratory sessions).
Jurisdictions: Benefit: Baltimore, Maryland, USA. Developer: United States. Operator: United States.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Conditions for replication: Signal power modulation remains stable (a transient reduction interrupted use of the fixed model). Laboratory setting with research staff.
Replication attempts: 1 follow-up entries. Benefit recurred: 1.
Later findings: An October 2025 medRxiv preprint concerns longitudinal gesture decoding in the same CortiCom trial. Its findings remain unverified.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Assessed benefits: People. Organizations. Society.
Depth: People: 2. Organizations: Not assessed. Society: Not assessed.
Reasons: People: Functional text-based communication at 10.2 characters per minute in laboratory sessions totaling 626 minutes over 90 days. No home use. Organizations: No organizational outcome reported. Society: Single participant.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: The makers' claim

Kind: Not assessed
Related entries: Not assessed
Field: Health; assistive communication. People with ALS and impaired speech or movement.
Purpose: Restore text communication for a person with ALS through a single neural click command driving a switch-scanning speller
Dates: Date basis unassessed: 2024 (precision unassessed). Updated: 22 October 2026.
Stage: Not assessed
Version: SHA-256: b059f22bdf11…