Machine learning will mean more drug ads, and hopefully better outcomes, says ad-tech firm DeepIntent

If you’ve ever visited a doctor, you may find yourself receiving more ads for drugs in coming years.

Advertising by Big Pharma directly to consumers is a small portion of the total online advertising market but may increase as new advertising tools, some of them using machine learning, are employed by the drug companies.

“Pharma is about 18% of national GPD and only 3% of digital advertising, that’s pretty astonishing,” Christopher Paquette, CEO of New York-based DeepIntent Technologies, told ZDNet in a telephone interview. DeepIntent, founded just over four years ago, is part of publicly traded advertising technology firm Propel Media.

“There is a $20 billion opportunity to unlock this digital advertising,” said Paquette. With more and more marketing moving from so-called linear TV to streaming, “pharmaceutical companies start to need to find new ways to reach patients,” he said. “We are at this inflection point.”

Paquette, formerly head of data science at Memorial Sloan Kettering Cancer Center in New York, is quick to point out that DeepIntent is “not an ad-tech company.”

“We don’t do advertising for advertising’s sake,” he told ZDNet, “We do it to get info to patients to make more informed decisions.”

On Thursday, DeepIntent announced the availability of a new product for Big Pharma to close the online drug marketing gap, called Patient Modeled Audiences. The service is the first ever for DeepIntent in which it has helped clients reach individuals. Previously, the company focused on helping drug company sales reps connect to physicians. The product has been in beta for about six months and took about a year to develop. Some customers already have been running live ad campaigns with the technology.

DeepIntent counts seven of the top ten pharmaceutical companies as clients, said Paquette. He declined to name the clients.

What DeepIntent does is develop machine learning programs that match data on patients with data from consumer marketing firms. The idea is to find correlations between the two data sets that let DeepIntent’s clients direct their campaigns to people more likely to need the drug. 

People can’t be identified directly for marketing purposes for privacy reasons, so DeepIntent has developed a way to infer prospective customers based on what medical data says about consumer data.

“We can’t target type-II diabetics directly, but we can target men who are 45 years old from the south who watch late night television because we know that they over-index for type-II diabetes” as a group, said Paquette. 

How the company does it is a combination of existing machine learning technology and proprietary code developed in-house.

The main work of classifying individuals in order to target them for marketing is done with a well-established, open-source machine learning tool called XGBoost. Developed by Tianqi Chen, a Carnegie Mellon professor, XGBoost builds on several decades’ worth of a statistical approach called decision trees. A decision tree takes aspects of a person such as age or location and ranks those attributes in order of importance, such as “is male,”http://www.zdnet.com/”is from the south,”http://www.zdnet.com/”watches late night TV.” The value of those attributes are summed to come up with a total score for an individual. Thoudands of such demographic variables  can be compiled to score which person might be a good candidate for an ad campaign. 

Data from actual patients is used to develop a decision tree that picks out matching attributes in the consumer population whose identities are known. (People in the consumer database have over the years volunteered to be marketed to through opt-in procedures.)

In other words, the private healthcare database tells the machine learning program what to look for in the public data set as far as relevant marketing attributes.

Consider as an example the drug Vascepa, a treatment for cardiovascular disease made and sold by Amarin Pharmaceuticals PLC, a campaign for which was recently run on DeepIntent’s platform. By building a decision tree based on the patient data and the consumer data, DeepIntent infers the audience of people who are likely to have been diagnosed with cardiovascular disease. “Users who score high (meaning they exhibit the same type of demographic, occupational, and lifestyle similarities as actual cardiovascular patients)” will be sent an ad telling them that Vascepa has just gotten an FDA approval, Paquette explained. 

If users are later diagnosed with cardiovascular disease, Paquette said, when the doctor prescribes Vascepa, “they’re already familiar with its benefits and therefore more likely to start and adhere to the prescribed treatment.”

The non-medical demographic data is obtained from “really well known marketing databases,” said Paquette, offering Experian as an example. That consumer data needs to be compared against private health records, which are kept by companies that process medical claims and pharmacy claims. Paquette is not at liberty to disclose the names of those companies, but he said DeepIntent is able to work with the firms through an arrangement with Datavant, a San Francisco firm that offers services to Big Pharma.

Paquette, who trained in the field of bioengineering, describes his prior job, at Memorial Sloan Kettering, as “breaking down silos” of different patient information systems. In an odd twist, at DeepIntent, Paquette is in a sense navigating a necessary boundary between silos, the public and the private.

The claims processors to which Datavant connects DeepIntent can’t show DeepIntent any patient-specific records because of the U.S. Health Insurance Portability and Accountability Act of 1996, or HIPAA. 

That fact meant that Paquette and team had to set up a firewall, a secure computing environment through which they would send application programming interface calls to access the combined consumer and health data without directly accessing the data.

That airlock, or gateway, if you will, is maintained “within a leading cloud services provider, managed only by the data providers with physical and logical separation from DeepIntent by design for HIPAA compliance,” Paquette told ZDNet (emphasis Paquette’s).

Despite that secure gateway, there is a tricky issue regarding privacy and statistical surveys. It has been shown on numerous occasions that statistical models can leak identities of those being surveyed. 

In one widely noted example, researchers have shown how anonymized data in the U.S. Census can be reverse-engineered with software tools to infer the actual identity of U.S. citizens, as has been nicely explained by Columbia University Journalism School professor Mark Hansen.

Paquette is well aware of such leakage risk.  

“There’s a stat out there that 87% of Americans are unique just by looking at zip code, gender, and date of birth,” said Paquette.

“So when you’re working with 1,000s of attributes of an individual, there’s definitely significant risk to identify specific patients and their [medical] conditions based on that demographic data.”

That vulnerability “added a twist” to the engineering, said Paquette. To avert the potential leak of personal information from the anonymized healthcare data, DeepIntent has developed its own custom code for what is known as differential privacy. 

Differential privacy is a broad class of algorithmic approaches to inject noise into data to make it harder to reverse-engineer the data to obtain true identity. 

Both DeepIntent’s firewall procedure, and its approach to differential privacy, was assessed by a firm called Mirador Analytics Ltd. of Roxburghshire, U.K. which specializes in things such as making statistical disclosure assessments. The assessment report, which was sent to ZDNet, found that the dataset used by DeepIntent “provided only ‘very small’ risk of disclosure,” and “can, therefore, be considered as meeting HIPAA de-identification standards.” 

Asked by ZDNet if DeepIntent will open its differential privacy code to third-party scrutiny more broadly, Paquette replied that DeepIntent “are working on putting together a patient and healthcare provider advisory board to help guide our platform policies.”

“While we don’t have a current plan to open the differential privacy code to third party developers, it’s something we may consider for the future,” he added.

DeepIntent, Paquette argues, is “solving the Goldilocks problem,” reaching the right patients but not being so precise that its technology would risk memorizing actual individuals. 

That gets to the larger question of the common good. What is such marketing for? The economic benefits to clients, the drug companies, could be substantial if Paquette is right that the approach is far more efficient.  Analysis of DeepIntent’s audiences has been performed by outside firms such as Crossix Solutions that measure effectiveness of media. “We are able to reach on-target patients who are eligible to learn more about a given treatment at 60% more efficient, less cost than the other competitive approaches,” said Paquette.

“What they are looking at is front-end metrics,” said Paquette, referring to the drug companies. “How on-target are you? How efficient from a media cost standpoint are you?” 

“From that standpoint, we are outperforming.” (More information is available in a case study posted on the company’s Web site.)

More data on the exact return-on-investment, as far as whether it leads to sales, may come from clients themselves, down the road, said Paquette, and that may help DeepIntent further improve its product.

Cynics may scoff at increased drug marketing. To that, Paquette presents an ambitious argument that DeepIntent can do well by patients by making them better informed. 

“If you think about the core principles that we founded DeepIntent on, it was to get information to patients to make more informed decisions, that’s our North Star,” he said. 

The roadmap of the product, he said, includes future development of ways to measure how ads are leading to better patient outcomes. That could include raising awareness about something like a COVID-19 outbreak, or increasing awareness in the population about rare diseases.

“To do just targeting would be short-sighted,” said Paquette. 

“We want to actually take it one step further, to be able to say, to prove out, with a measurable, quantitative approach, Hey, this is actually the amount of change, the amount of improvement in patient outcomes that can be attributed to the media and the influence that you’ve had through our  platform.”