Meghna Roy Chowdhury

MyoText: Hierarchical sEMG-to-Text Decoding via Finger-Intent Modeling

The problem

Wrist-worn sEMG could make keyboard-free typing possible, but the signal records which finger moved, not which key it pressed. Decoding characters directly forces a model to separate keys typed by the same finger (r, t and g are all left index), which sEMG barely distinguishes.

How MyoText works

MyoText is a hierarchical framework that adds finger intent as a step between muscle signal and text. It is trained and tested on Meta's emg2qwerty dataset (108 participants, over 346 hours of 32-channel wrist sEMG).

1. Predict the finger: a CNN–BiLSTM–Attention classifier maps each 2 s window of 32-channel sEMG to one of eight fingers.
2. Constrain the letter pool: QWERTY touch-typing rules turn each finger into a small letter pool, and dictionary filtering keeps real words. For “this”, 216 letter combinations shrink to 3 candidates.
3. Decode the sentence: a fine-tuned T5 transformer picks one word per position using sentence context.

What it shows

• Finger is the better target: 85.4% finger accuracy, against 51.2% when the same model predicts the 26 letters directly.
• New users without calibration: 5.4% CER and 6.5% WER on 4 held-out users. The standard CTC + 6-gram baseline, retrained on the same 30 users, reaches 64% CER. Training on 65 users lowers CER further, to 2.73%.
• Tolerates non-standard typing: letting letters map to neighboring fingers lowers CER to 3.7% instead of raising it.

What's next

All results so far are offline on emg2qwerty. Next:

• A real-time, human-in-the-loop study on a physical wrist-worn device.
• Lower latency: the 2 s window adds 1 s of look-ahead, so a sentence is ready 1.155 s after the last keystroke. A shorter window could cut that.
• Other languages and typing styles (for example hunt-and-peck or thumb typing), by swapping the key mapping and decoder while keeping the finger classifier.

Interpretable Hierarchical Machine Learning for Mental Health

The problem

Phones leave behind huge trails of behavioral data, which could say a lot about mental health. But that data is noisy, high-dimensional, and varies wildly from person to person, so ordinary models struggle with it.

What I built

I-HOPE is a hierarchical model that turns messy raw behavior into readable interaction patterns. Careful feature reduction trims the data without throwing away what predicts well, so the results stay both personal and interpretable.

How it turned out

• 91% accuracy on student mental-health outcomes, up from a 70% baseline.
• A clear path from raw behavior to meaningful indicators, instead of a black box.
• Per-student explanations for every prediction, using SHAP.

Distributed Neural Networks for IoB

The problem

Smarter wearables need more compute, and compute drains the battery fast. The question was how to run capable AI on a device that has almost no energy to spare.

What I built

DistNN splits a neural network between the tiny wearable and a more powerful nearby hub, like your phone, and finds the best point to make that cut. Custom low-precision hardware handles the wearable's share at very low cost.

How it turned out

• About 1000x more energy-efficient than running the same AI on a GPU.
• Roughly 11x less power than recent state-of-the-art wearable AI chips.
• Still real-time (30 fps) and accurate (0.9 SSIM on image reconstruction).
• The wearable's share of the model runs at microwatt-level power.

Self-Supervised Learning with Squeeze-and-Excitation Networks for EEG

The problem

EEG matters for brain-computer interfaces and for diagnosing neurological conditions, but real-world recordings are messy, usually unlabeled, and expensive to process on low-power devices.

What I built

SSL-SE-EEG pairs self-supervised learning, which learns from unlabeled EEG, with lightweight squeeze-and-excitation networks. Turning the raw signals into structured 2D images makes them far easier for the models to read.

How it turned out

• 91% on MindBigData and 85% on TUH-AB seizure detection.
• Much less dependence on expensive labeled data, thanks to the self-supervised step.
• More robust to the noise and artifacts that plague real EEG.
• Low enough overhead to suit real-time, low-power BCI.

Low Power Wearable ECG via Human Body Communication

The idea

Continuous ECG usually means a bulky device you keep recharging. FlexECG skips the battery entirely: it's a flexible patch that harvests heat from your body and sends the ECG through your body itself using human body communication (HBC).

The hard part

Body heat provides very little power, and not steadily. Early versions kept browning out mid-reading, so most of the work went into getting the system to run reliably inside that budget.

How it turned out

• A custom HBC transmit-and-decode path for ECG signals.
• Settles at about 20µW of power.
• Tested across multiple human subjects.
• Roughly 4x lower power than comparable state-of-the-art systems.

Electric Field-Based Intrathecal Drug Delivery (IDD)

The problem

Intrathecal drug delivery sends medication straight to the central nervous system for long-term pain relief. The hard part is being sure the catheter actually sits in the cerebrospinal fluid (CSF), not an adjacent layer, without adding cost or complexity.

The approach

CSF conducts electricity better than the tissue around it at certain frequencies (tens of MHz). This system uses a galvanic transmitter and a capacitive receiver at electroquasistatic (EQS) frequencies to read that difference and infer where the catheter tip is.

What the simulations showed

• Modeled in Ansys HFSS on the realistic “Neva” human-body model.
• A clear difference in field strength when the transmitter sits in the CSF versus neighboring layers.
• A way to verify catheter placement for more precise delivery.
• As far as we know, the first electric-field approach proposed for IDD.

AI for Safer Surgery: Predicting Anesthesia Depth

The problem

During surgery, knowing exactly how deeply someone is anesthetized keeps them safe and comfortable, and the usual monitors have limits. The question was whether AI could read it from signals you're already recording.

The approach

Using ECG and PPG recorded from real surgical patients, I trained deep models (CNNs and LSTMs) to estimate anesthetic depth in real time.

How it turned out

• Processed messy surgical ECG/PPG in MATLAB and built the models in Python (TensorFlow/Keras).
• About 82% accuracy predicting clinical anesthesia depth from ECG features alone.
• Published in the BSPC journal.
• This was my first real paper, from my time at Yuan Ze in Taiwan, and the one that got me hooked on health AI.