[TheNetherWatcher]

When a signal has more than one sample per frame, smooth_frames averages each frame in digital units. It doesn't check for the invalid sample value, so an invalid sample is averaged with the others and gives a large value that looks valid:

import numpy as np, wfdb

# fmt 16, 2 samples per frame, gain 200. -32768 is the invalid sample value.
np.array([0, 1, -1, -2, -32768, 100], dtype="<i2").tofile("r.dat")
open("r.hea", "w").write("r 1 100 3\nr.dat 16x2 200/mV 16 0 0 0 0 sig\n")

wfdb.rdrecord("r", smooth_frames=False).e_p_signal[0]
# [ 0.     0.005 -0.005 -0.01     nan  0.5  ]
wfdb.rdrecord("r").p_signal[:, 0]
# [ 0.    -0.005 -81.67]   (last frame should be NaN)

The WFDB library returns an invalid value for any frame that contains an invalid sample (meansamp() in lib/signal.c), so this does the same: after averaging a chunk of frames in digital units, frames with an invalid sample are set to the invalid value for the channel's format. That then becomes NaN when converted to physical units. Physical smoothing (sigtype="physical") already gave NaN, since NaN propagates through the sum.

This doesn't change the rounding of valid frames, which is #335.

Tests: a fmt 16 record with two samples per frame and one invalid sample reads back as [3, -32768, 7] (digital) and [0.015, nan, 0.035] (physical). This fails on main. pytest -m "not network" passes (97 tests), and the docs build passes.

[tompollard]

Thanks, this looks great!

I compared it against rdsamp from the WFDB C library (frame-averaged output) on the multi-frequency sample records that have invalid samples:

• wave_4: rdsamp gives 3659 invalid frames. With this PR the Python reader gives the same 3659, at the same positions. On main it gave 7.
• 041s02: one invalid frame in both, at the same position. On main there were none.

All other multi-frequency records in sample-data read exactly as before.