4 ms·

In scientific research, data mining as it's apparently often practiced in the private sector isn't a thing - there's study design, there's controls, there's a w

In scientific research, data mining as it's apparently often practiced in the private sector isn't a thing - there's study design, there's controls, there's a whole system of how to go about collecting data that's been designed in advance so that datasets are collected that can be used to answer the questions of interest. Garbage in, garbage out, that's worth remembering.

crabbone 4y ago
Here's an example from my wife: they are working in cooperation with some hospital. They were tasked with extracting breathing information from the EKGs. (This could be helpful in detecting when to take patients off the ventilator). Worked for half a year with some previously acquired data. The hospital part of the team aren't particularly tech-savvy and didn't participate in the research as much. But once they actually had to produce current EKG data, it turned out that the sampling frequency of their equipment is half of the minimum required in order to detect breathing. Now the research is worth nothing. Another similar story where I got tangentially involved: my wife had to reproduce a notebook of some guy who wrote a thesis on detecting something in EEG signals. After a lot of digging and research into that guy's work, I discovered that the format they used to acquire EEG signal in could have either 32 or 16 bits per "event", the event streams were not contiguous, but per sampling per electrode attached to the scalp. I.e. physically, they were written as event1_electorde1, event1_electrode2, ... event1_electrodeN, ... event2_electrode1, ... etc. The original research author didn't understand this aspect, and the data he used was generated by different EEG machines in different format. He used some framework function that tried to guess how many bits per signal were used, and when it failed to guess, it'd print a warning. He simply silenced the warnings. And that's how he ran his ML models, and that's how he submitted his thesis. He has a graduate degree now and had moved to a different country. His work is worthless.

crabbone 4y ago

Here's an example from my wife: they are working in cooperation with some hospital. They were tasked with extracting breathing information from the EKGs. (This could be helpful in detecting when to take patients off the ventilator). Worked for half a year with some previously acquired data. The hospital part of the team aren't particularly tech-savvy and didn't participate in the research as much. But once they actually had to produce current EKG data, it turned out that the sampling frequency of their equipment is half of the minimum required in order to detect breathing. Now the research is worth nothing. Another similar story where I got tangentially involved: my wife had to reproduce a notebook of some guy who wrote a thesis on detecting something in EEG signals. After a lot of digging and research into that guy's work, I discovered that the format they used to acquire EEG signal in could have either 32 or 16 bits per "event", the event streams were not contiguous, but per sampling per electrode attached to the scalp. I.e. physically, they were written as event1_electorde1, event1_electrode2, ... event1_electrodeN, ... event2_electrode1, ... etc. The original research author didn't understand this aspect, and the data he used was generated by different EEG machines in different format. He used some framework function that tried to guess how many bits per signal were used, and when it failed to guess, it'd print a warning. He simply silenced the warnings. And that's how he ran his ML models, and that's how he submitted his thesis. He has a graduate degree now and had moved to a different country. His work is worthless.