Big data’s dirty problem
Karim Keshavjee, a Toronto physician and digital health consultant, crunches mountains of data from 500 doctors to figure out how to improve patient treatment. But it’s a frustrating slog to get a computer to decipher all the misspellings, abbreviations, and notes written in unintelligible medical shorthand.
For example, “smoking information is very hard to parse,” Keshavjee said. “If you read the records, you understand right away what the doctor meant. But good luck trying to make a computer understand. There’s ‘never smoked’ and ‘smoking = 0.’ How many cigarettes does a patient smoke? That’s impossible to figure out.”
The hype around slicing and dicing massive amounts of data, or big data, makes it sound so easy: Just plug a library’s worth of information into a computer and wait for valuable insights to pour out about how to speed up an auto assembly line, get online shoppers to buy more sneakers, or fight cancer. The reality is much more complicated. Data is inevitably “dirty” thanks to obsolete, inaccurate, and missing information. Cleaning it up is an increasingly important and overlooked job that can help prevent costly mistakes.
Although techniques are improving all the time, scrubbing data can only accomplish so much. Even when dealing with a relatively tidy set of information, getting useful results can be arduous and time-consuming.
“I tell my clients that the world is messy and dirty,” said Josh Sullivan, a vice president at business consulting firm Booz Allen who handles data crunching for clients. “There are no clean data sets.”
Data analysts start by looking for information that’s out of the norm. Because the volume of data is so huge, they typically hand the job over to software that automatically sifts through numbers and text to look for anything unusual that needs further review. Over time, computers can improve their accuracy in spotting what’s belongs and what doesn’t. They can also better understand what words and phrases mean by clustering similar examples together and then grading their interpretations for accuracy.
“The approach is easy and straightforward, but training your models can take weeks and weeks,” Sullivan said.
A constellation of companies offer software and services for cleaning data. They range from technology giants like IBM (IBM) and SAP (SAP) to big data and analytics specialists like Cloudera and Talend Open Studio. A legion of start-ups are also trying to get a toehold as data janitors including Trifacta, Tamr, and Paxata.
Healthcare, with all its dirty data, is one of the toughest industries for big data technology. Electronic health records make medical information increasingly easy to dump into computers, but there’s still a lot room for improvement before researchers, pharmaceutical companies and hospital business analysts can slice and dice all the information they want.
Keshavjee, the doctor and CEO of InfoClin, a health data consulting firm, spends his days trying to tease out ways to improve patient treatment by sifting through tens of thousands of electronic medical records. Obstacles pop up all the time.
Many doctors neglect to note a patients blood pressure in their medical records, something that no amount of data cleaning can fix. Simply determining what ails patients—based on what’s in their files—is surprisingly difficult for computers. Doctors may enter the proper code for diabetes without clearly indicating whether it’s the patient who has the disease or a family member. Or they may just enter “insulin” without mentioning the underlying diagnosis because, to them, it’s obvious.
Physicians also use a lot of idiosyncratic shorthand for medications, illnesses and basic patient details. Deciphering it takes a lot of head scratching for humans and is nearly impossible for a computer. For example, Keshavjee came across one doctor who used the abbreviation”gpa.” Only after coming across a variation, “gma,” did he finally solve the puzzle—they were shorthand for “grandpa” and “grandma.”
“It took a while to figure that one out,” he said.
Ultimately, Keshavjee said one of the only ways to solve the problem of dirty data in medical records is “data discipline.” Doctors need to be trained to enter information correctly so that cleaning up after them is less of a chore. Incorporating something like Google’s helpful tool that suggests how to spell words as users type them would be a great addition for electronic medical records, he said. Computers can learn to pick out spelling errors, but minimizing the need is a step in the right direction.
Another of Keshavjee’s suggestions is to create medical records with more standardized fields. A computer would then know where to look for specific information, reducing the chance of error. Of course, doing so is not as easy as it sounds because many patients suffer from multiple illnesses, he said. A standard form would have to be flexible enough to take such complications into account.
Still, doctors would need to be able to jot down more free-form electronic notes that could never fit in a small box. Nuance like why a patient fell, for example, and not just the injury suffered, is critical for research. But software is hit and miss in understanding free-form writing without context. Humans searching by keyword may do a better job, but they still inevitably miss many relevant records.
Of course, in some cases, what appears to be dirty data, really isn’t. Sullivan, from Booz Allen, gave the example the time his team was analyzing demographic information about customers for a luxury hotel chain and came across data showing that teens from a wealthy Middle Eastern country were frequent guests.
“There were a whole group of 17 year-olds staying at the properties worldwide,’ Sullivan said. “We thought, ‘That can’t be true.’ ”
But after some digging, they found that the information was, in fact, correct. The hotel had legions of young customers that it didn’t even realize were there, and had never done anything to market to them. All guests under 22 were automatically logged as “low-income” in the company’s computers. Hotel executives had never considered the possibility of teens with deep pockets.
“I think it’s harder to build models if you don’t have outliers,” Sullivan said.
Even when data is clearly dirty, it can sometimes be put to good use. Take the example, again, of Google’s spelling suggestion technology. It automatically recognizes misspelled words and offers alternative spellings. It’s only possible because Google (GOOG) has collected millions and perhaps billions of misspelled queries over the years. Instead of garbage, the dirty data is an opportunity.
Ultimately, humans, and not machines, draw conclusions from the data they crunch. Computers can sort through millions of documents, but they can’t interpret the findings. Cleaning data is just one of step in a long trial and error process to get to that point. Big data, for all its hype about its ability to lift business profits and help humanity, is a big headache.
“The idea of failure is completely different in data science,” Sullivan said. “If you they don’t fail 10 or 12 times a day to get to where they should be, they’re not doing it right.”