wfdb

[Litsardos]

wfdb package for xqrs_detect

[ndonkoHenri]

wfdb needs no mobile-forge recipe — it already works on both platforms straight from PyPI.

It is pure Python (wfdb-4.3.1-py3-none-any.whl), and every native package in its dependency tree already has mobile wheels: numpy, scipy, pandas, matplotlib, contourpy, kiwisolver, pillow from pypi.flet.dev, and aiohttp / charset-normalizer from PyPI's own Android/iOS wheels. fsspec and requests are pure Python.

Example

requires-python = ">=3.12"
dependencies = [
    "flet",
    "wfdb==4.3.1",
]

Add the below only if your app plots (wfdb.plot_items / plot_wfdb, or your own matplotlib import):

[tool.flet.android]
extract_packages = ["matplotlib"]

matplotlib reads mpl-data/matplotlibrc through a real __file__ path, and Flet 0.86 ships site-packages as a compressed sitepackages.zip on Android — without that line you get NotADirectoryError. It is an app-side setting; the recipe's own extract_packages does not propagate to consumers.

Example: Detecting beats

import asyncio
import functools

import flet as ft
import numpy as np
from wfdb.processing import calc_mean_hr, calc_rr, xqrs_detect

FS = 250

def main(page: ft.Page):
    result = ft.Text()

    async def detect(e):
        sig = np.asarray(my_ecg_samples, dtype=np.float64)   # one channel, in mV
        # xqrs is CPU-bound — keep it off the UI thread.
        qrs = await asyncio.to_thread(
            functools.partial(xqrs_detect, sig=sig, fs=FS, verbose=False)
        )
        rr = calc_rr(qrs, fs=FS, rr_units="seconds")
        hr = calc_mean_hr(rr, fs=FS, rr_units="seconds")
        result.value = f"{len(qrs)} beats — {hr:.1f} bpm"
        page.update()

    page.add(ft.FilledButton("Detect QRS", on_click=detect), result)

ft.run(main)

import wfdb.processing pulls in scipy and can take a noticeable moment on a phone — do it off the first frame, and keep xqrs_detect off the UI thread as above. From a plain (non-async) handler the equivalent is page.run_thread(...).

Caveats worth knowing

1. FLAC-format records don't work. wfdb reads WFDB formats 508/516/524 through soundfile, and soundfile's pure wheel carries no libsndfile, so import soundfile raises OSError on device. soundfile is imported lazily (wfdb/io/_signal.py:1886, :2482) so it costs you nothing until you open such a record. Common ECG datasets are unaffected — mitdb is format 212, PTB-XL is 16 — but MIMIC-IV Waveform (mimic4wdb) is 516x4 and would need a flet-libsndfile recipe. Happy to add one if someone needs it.

2. requires-python picks the highest supported Python, not the lowest. >=3.12 resolves to 3.14. Write "==3.12.*" to pin.

3. 32-bit x86 emulators. aiohttp/pandas are only published for android_24_x86 on cp312, so a 32-bit x86 AVD won't resolve on 3.13/3.14. Use an x86_64 or arm64 emulator.

4. If anything imports matplotlib, set MPLCONFIGDIR. On Android matplotlib's default config dir resolves to /data/.matplotlib, which isn't writable, so it rebuilds its font cache into a purgeable temp dir on every cold start:

os.environ.setdefault(
    "MPLCONFIGDIR",
    os.path.join(os.getenv("FLET_APP_STORAGE_DATA", "."), "matplotlib"),
)

Put it above the first matplotlib import.

5. Size. scipy (89 MB), pandas (46 MB), numpy (23 MB) and matplotlib (22 MB) dominate: ~180 MB of site-packages per ABI, and the little demo app I built lands at 177 MB APK for a single --arch arm64-v8a. matplotlib is dead weight for a detect-only app but is a hard Requires-Dist of wfdb, so it ships regardless. Build one ABI at a time, and consider [tool.flet.cleanup] package_files = ["scipy/**/tests", "matplotlib/tests"].

---

So: nothing to build on our side — add wfdb to your dependencies and it works. Let know if you face issues or need assistance.