I agree with Raghu in that a distributed system is a better way to go but to directly address your problem, why do you no go away from the idea of using interrupts from the 8574's and just regularly poll the 8574's? How fast do you want to poll the switches? 10mS? That might be a good start. If you have i2C hardware in your micro, the overhead will be minimal, and because you sample at a time interval, the debounce code should be easier. If each switch bit is shifted into a 8 bit byte, it is just a matter of testing for a number of 1 bits in a sequence - if you sample at 10mS and the last 3 samples were all 1's then the input has been stable for 30mS - if it was 101 then we know the switch is still bouncing.