Inter-channel timing in the continuous acquisition modes

[h-mayorquin]

Hi,

Heberto again, from the double sampling thread. I am writing a pyPhotometry interface for neuroconv, and timing is the part of an NWB file I am most careful about, since it is what everything else in a session gets aligned against. What follows is almost certainly too small to be behaviourally relevant, but I would rather not assume that if I do not have to.

In the continuous modes continuous_ISR reads the two ADCs one after the other, and MicroPython documents read_timed as blocking until its buffer is full:
https://github.com/pyPhotometry/code/blob/v1.1/uPy/photometry_upy.py#L146-L159

But import_ppd returns one time vector for every analog signal:
https://github.com/pyPhotometry/code/blob/v1.1/tools/data_import.py#L97

With a 64 sample buffer and oversampling_rate at 300 kHz, analog 2 looks like it starts at least 213 us after analog 1, plus whatever the lines between the two reads cost. I can bound it the other way too, since the whole interrupt has to fit inside one sample period, so at the 1 kHz ceiling it is somewhere under about 787 us. What I cannot do is turn that into a number, as neither constant is in the header.

So for now I write the continuous signals on the shared timebase and say in the file that they are not simultaneous. Is that shared time vector meant as "the offset is small enough to ignore", or should a reader be offsetting the second channel by a figure you have measured? Is this hardware detail documented somewhere?

[ThomasAkam]

Hi @h-mayorquin,

Thanks for your work on integrating pyPhotometry data into neuroconv.

The channel 1 and channel 2 signals are indeed read at a temporal offset from each other. We have never taken this into account when analysing data generated by pyPhotometry, as typically the signals are lowpass filtered to <10Hz early in the preprocessing pipeline so the time offset between the samples of the two signals is small relative to the timescale of signal variation.

Additionally, as we typically synchronise the photometry signal with other data streams using sync pulses output by the behavioural hardware and recorded by pyPhotometry at its sampling rate (see synchronisation docs here), the time resolution of when sync pulses occued is limited by the photometry signal sampling rate, so we have not worried about the precision of sample times below this. It is possible to get more precise synchronisation by using the functionallity to output sync pulses from pyPhotometry as then the sync pulse output is activated immediately before the analog 1 signal is read.

In the time division acqusition modes (which are recommended for recording brain signals) where the LEDs for the two channels are pulsed in anti-phase with each other, the channel 1 and channel 2 samples will be seperated in time by half the sampling period. In the continuous acquisition mode the offset will, as you say above, be deteremined by the delay between the ADC1 and ADC2 channel reads. I just measured the delay between the channel 1 and channel 2 reads in continuous mode, using an oscilloscope and a modified version of the board firmware shown below that turns on the LED for each channel just before the ADC reads, and got a mean delay between the two reads of 393us, with a standard deviation of 9us.

    @micropython.native
    def continuous_ISR(self, t):
        # Modified version of Interrupt service routine for 2 color continous acquisition mode which turns on LEDs just before ADC reads.
        if self.sync_out:
            self.sync_pulse_update()
        self.LED1.write(self.LED_1_value)
        self.ADC1.read_timed(self.ovs_buffer, self.ovs_timer)  # Read sample of analog 1.
        self.sample = sum(self.ovs_buffer) >> 3
        DI1_value = self.sync_pulse_state if self.sync_out else self.DI1.value()
        self.sample_buffers[self.write_buf][self.write_ind] = (self.sample << 1) | DI1_value
        self.write_ind += 1
        self.LED2.write(self.LED_2_value)
        self.ADC2.read_timed(self.ovs_buffer, self.ovs_timer)  # Read sample of analog 2.
        self.sample = sum(self.ovs_buffer) >> 3
        self.sample_buffers[self.write_buf][self.write_ind] = (self.sample << 1) | self.DI2.value()
        self.LED1.write(0)
        self.LED2.write(0)
        # Update write index and switch buffers if full.
        self.write_ind = (self.write_ind + 1) % self.buffer_size
        if self.write_ind == 0:  # Buffer full, switch buffers.
            self.write_buf = 1 - self.write_buf
            self.send_buf = 1 - self.send_buf
            self.buffer_ready = True

[h-mayorquin]

Wow, so this is more or less what we expected, thanks for measuring it directly.

I am adding this to our documentation of the format so the users are informed but I agree with you that the adjustment is small enough that applying it does not make sense. True alignment needs an external signal to set the times right and this is more or
less what we recommend as well:
https://github.com/catalystneuro/neuroconv/blob/new_timing_model/docs/user_guide/temporal_alignment.rst

Thanks for the link to your docs, I have added this to my references.

Feel free to close this and thanks again for your quick response.