Epiread Talk to the team

Automated IED detection for clinical EEG

Spike detection for the EEGs nobody has time to read

Epiread screens hours of long term EEG in seconds and flags interictal epileptiform discharges at 60% recall with one false positive per hour. That is a rate a clinician can actually work with.

Talk to the team Built on clinical EEG from a tertiary epilepsy centre
Screening Longitudinal bipolar 8 ch 128 Hz
Flagged 0 Artifacts passed 0 Threshold p ≥ 0.78

Simulated EEG for illustration. Discharges in the left temporal chain, with phase reversal at T3, are flagged. A frontal eye blink is not.

The problem

The bottleneck is not the recording. It is who reads it, and for how long.

Epilepsy diagnosis depends on catching interictal spikes in EEG. The recordings exist. The specialist hours to read them do not.

50M
people live with epilepsy worldwide
WHO
11%
sensitivity of a first routine EEG after a first unprovoked seizure
Hernandez-Ronquillo et al., 2023
0.49κ
agreement between expert readers on individual discharges
Jing et al., 2020

Not enough readers

Specialist neurological care in India sits in the major cities. District and tier 2 hospitals often have EEG machines but no one trained to interpret the recordings.

The WHO African region has a median of 0.1 neurological specialists per 100,000 people. High income countries have 7.1, roughly seventy times as many.

WHO and World Federation of Neurology, Neurology Atlas, 2017

Not enough time

A first routine EEG caught IEDs or seizures in only 11% of adults after a first unprovoked seizure. A 24 hour ambulatory recording reached 72%, but that is 24 hours of signal someone has to read.

Hernandez-Ronquillo et al., Neurology: Clinical Practice, 2023

Interictal activity follows multiday cycles, most often 20 to 30 days long. A single short EEG is a snapshot at a random point in that cycle. Long term monitoring captures what short recordings miss by design.

Baud et al., Nature Communications, 2018

Even experts disagree. Eight expert readers scoring 13,262 candidate waveforms agreed on individual discharges with a κ of only 0.49. Artifacts mimic spikes, and morphology varies widely across patients.

Jing et al., JAMA Neurology, 2020

The goal is not to replace the epileptologist. It is to make sure their hours go to the recordings that matter.

Our approach

A clinician reviews minutes instead of hours

Epiread screens the full recording automatically and surfaces only the segments most likely to contain spikes. The operating point is chosen for the reading room, not the leaderboard.

60%
recall at one false positive per hour
1/hr
false positive rate at that recall
8,000+
clinical EEGs in the data pipeline
2,000+
24 hour recordings collected

A low false alarm rate is the point

A detector that raises dozens of false alarms per hour creates more work than it saves. At one per hour, every flag is worth a look, and the output fits a real clinical workflow without alert fatigue.

Trained on clinical data, not benchmarks

All training data comes from a tertiary epilepsy centre: real patients, real noise, real artifacts. No clean room recordings that fall apart on ward data.

Every detection is explainable

The model reports which channels and temporal features drove each detection, so the reviewer knows where to look and can override with confidence.

Built for the settings that need it most

Compact enough to run in resource constrained hospitals. The epilepsy burden is highest where specialist access is lowest, and the model is sized for that.

How this differs from the FDA cleared standard

MeasurePersyst P15Epiread
EvidenceCook et al., 2026: 484 routine EEGs at a tertiary centreInternal validation on clinical EEG
Unit of analysisWhole recordingIndividual discharge
Sensitivity81.1%60%
False alarmsFlagged 40.1% of recordings with no IEDs1 per hour of recording
Explanation per detectionNot assessed in the studyChannels and temporal features behind each flag

The two columns come from different studies and different units of analysis, so this is not a head to head comparison. In Cook et al., only 19.9% of the recordings Persyst flagged were confirmed by the clinical report. Epiread is tuned the other way: lower peak sensitivity in exchange for a false alarm rate a reader can work with.

The value isn't catching every spike. It's making sure a 24-hour recording gets reviewed at all, in a center that otherwise wouldn't have the bandwidth.

Dr. Neeraj BahetiComprehensive Epilepsy Care Centre, CIIMS Nagpur

How it works

From a raw recording to a short review list

Every recording goes through the same three stages before a clinician sees it.

  1. 01

    Load the recording

    Recordings arrive as EDF or FIF from any 10–20 setup. Channels are mapped to the 19 standard electrodes and rebuilt into bipolar and average reference montages.

    EDF · FIF · 19 electrodes
  2. 02

    Screen every window

    The signal is bandpass filtered, notched for mains hum and resampled. The detector scores each window and keeps only those above its operating threshold.

    0.5 to 45 Hz · 50/60 Hz notch · 128 Hz
  3. 03

    Review what matters

    Flagged segments open in a paging viewer next to the original annotations. The epileptologist confirms or rejects each one, and the decision stays clinical.

    Flag at p ≥ 0.78 · 5 to 60 s pages

Team

Clinical access and signal processing, in one room

Epiread pairs direct access to clinical EEG and expert annotation with the machine learning to build on it.

Dr. Neeraj Baheti

Chief Medical Officer and Epileptologist

Founded Central India's first epilepsy surgery program at CIIMS Nagpur. Convenor of the Indian Epilepsy Society's EEG Task Force. His clinical network, with more than 700 professionals trained each year across international programs, is the data access and annotation backbone of Epiread.

DM Neurology, SCTIMSTFellowship in EpilepsyMBBS & MD Gold Medalist
45+peer reviewed publications
27h index, 950+ citations
700+professionals trained yearly

Samarth Adatia

Cofounder and AI Lead

B.Tech in Electrical Engineering and M.Tech in Signal Processing and Machine Learning from IIT Kharagpur. ML researcher at Stanford Medicine. Designed and built the full pipeline: data ingestion, preprocessing, model architecture, training infrastructure and clinical validation tooling.

IIT Kharagpur, EE + SP&MLStanford Medicine

Questions

What people ask first

Does Epiread replace the epileptologist?

No. It decides what gets looked at first. Every flag is reviewed by a clinician, who confirms or rejects it, and the diagnosis stays with them.

Is it approved for clinical use?

Not yet. Epiread is in development and is being prepared for review by CDSCO as Software as a Medical Device. Until then it is for research and evaluation only.

Which recordings does it work with?

EDF and MNE FIFF recordings with any channel naming. Channels are mapped to the 19 electrodes of the 10–20 system, and recordings that are already montaged run directly.

What hardware does it need?

Ordinary CPU hardware. The model is small enough to run without a GPU, which matters for the hospitals that need it most.

Who built it?

An epileptologist and a signal processing engineer in India, working from clinical EEG recorded at a tertiary epilepsy centre and annotated by specialists.

How do I get in touch?

Write to hello@epireadneuro.com. We read every message and reply from the team directly.

Get in touch

Running an EEG service or an epilepsy programme?

We would like to hear how your team reviews long term recordings today, and where the hours go.

Talk to the team