When writing a record (header + signal file) using Record.wrsamp, the signal data may either be single-frequency (d_signal) or multi-frequency (e_d_signal).

If expanded=True, then wrsamp takes the data from e_d_signal (ignoring d_signal) and writes one or more multi-frequency signal files. In this case the record must also have a samps_per_frame attribute, and this attribute is written to the header file.

If expanded=False (default), then wrsamp takes the data from d_signal (ignoring e_d_signal) and writes one or more single-frequency signal files. In this case the samps_per_frame attribute is still written to the header file, which is completely incorrect.

Why would you have a Record that is single-frequency yet has samps_per_frame not equal to 1? Because that's what rdrecord does by default when you read a multi-frequency record.

So:

>>> import wfdb
>>> r = wfdb.rdrecord('43700001', pn_dir='mimicdb/437', sampto=10, physical=0)
>>> r.wrsamp()
>>> wfdb.rdrecord('43700001')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/benjamin/wfdb-python/wfdb/io/record.py", line 3542, in rdrecord
    return_res=return_res)
  File "/home/benjamin/wfdb-python/wfdb/io/_signal.py", line 1010, in _rd_segment
    sampfrom, sampto, smooth_frames)[:, r_w_channel[fn]]
  File "/home/benjamin/wfdb-python/wfdb/io/_signal.py", line 1151, in _rd_dat_signals
    sig_data = _blocks_to_samples(sig_data, total_process_samples, fmt)
  File "/home/benjamin/wfdb-python/wfdb/io/_signal.py", line 1460, in _blocks_to_samples
    sig[0::2] = sig_data[0::3] + 256 * np.bitwise_and(sig_data[1::3], 0x0f)
ValueError: could not broadcast input array from shape (30) into shape (75)

(Arguably it's wrong for rdrecord to set samps_per_frame when smooth_frames=True, but changing that behavior would be hard to reconcile with rdheader, and would prevent implementing automatic frame-smoothing as I suggested in #313 (comment))