Amazon is famous for targeted marketing that approaches customers based on their unique needs. Like other successful businesses, such as Netflix, the company taps into machine learning, which uses customer data to understand their behavior.
Hospitals and medical centers rely on marketing too, investing heavily in direct-to-patient outreach to urge at-risk people to get regular screenings. Johns Hopkins Hospital's cancer center, for example, uses emails, letters, seminars and community events to encourage patients to get screened for potential cancer. The high cost of cancer treatment makes this effort worth it: research shows regular screenings help with early detection, leading to more cost-effective treatments and better prognoses.
But hospitals can – and must – improve their outcomes much more, by melding this essential outreach with individually tailored communications based on machine-learning insights.
In an award-winning paper, Rice Business Professor Vikas Mittal and colleagues developed new algorithms indicating that targeted, personalized outreach can increase screenings among at-risk patients. "Outreach marketing" – including sending informational letters and talking to patients about potential barriers to screening – was indeed a powerful motivator for patients to get screened, ultimately lowering health care costs for patient and hospital. But patients with different characteristics, Mittal's team found, responded differently to marketing interventions. When it came to marketing campaigns for cancer screening prevention, one-size-fits-all outreach efforts were neither effective nor economical. Personalized marketing works better for preventing cancer.
To conduct their research, the researchers randomly divided 1,800 patients at UT Southwestern Medical System at risk for hepatocellular carcinoma – the most common type of primary liver cancer – into three groups – usual care, outreach alone, and patient navigation, which includes help such as follow-up calls, motivational messages and assistance spotting specific barriers. They followed each group to see if patients scheduled an MRI or CT scan within six months, from 6-12 months and from 12-18 months.
The first group was asked to receive a screening during their doctors' visits and wasn't contacted after that. The second group received a one-page letter in the mail, then staff called patients who didn't schedule a screening. The third group receiving patient navigation got the same treatment as the second group supplemented with phone calls designed to identify potential barriers, which they used to give customized motivational messages encourage coming in for a screening.
The researchers used patient data from medical records, including patients' age, gender, ethnicity, income, commute time, health status, how often they received healthcare services, whether or not they had insurance and how populated their neighborhoods were.
Following traditional methods, Mittal's team found that the patients who got a letter and call were 10-20% more likely to complete a screening, while those who got the customized motivational messages were 13-24% more likely to schedule their screening. But this is where traditional medical research stops, without asking a crucial question: Within each group, such as those of the 600 patients receiving patient navigation, could screening rates differ based on patients' individual characteristics?
In past research, everyone receiving the same stimulus is presumed to respond the same way. There was no statistical technique to separately estimate the responsiveness of patients with different characteristics. Mittal's team solved this problem by using a machine learning technique called causal forests.
By using "causal forests" to quantify how each of the three marketing approaches could be applied to different patients, Mittal's team found, improved returns on the traditional approach by a remarkable 74-96% – or by $1.6 million to $2 million.
Using traditional methods, physicians would have concluded that every patient should get patient navigation because it was a more intensive marketing approach. The causal forest method showed otherwise: there are small groups of patients with unique characteristics who respond best to specific types of overtures. Minority women in good health who had insurance, visited the doctor often and lived close to clinics in more populated neighborhoods responded especially well to all three types of outreach interventions. Younger patients with long commutes who live in neighborhoods with more public insurance coverage embraced the second type of intervention, outreach alone. And older patients in higher-income neighborhoods favored the patient-navigation approach.
The stakes for common marketing practices like "AB testing" could not be higher. In AB testing, marketers run randomized experiments such as showing ads to some people and not to others. If those seeing an ad, on average, buy more, the conclusion is to blanket the market with ads. But AB testing ignores the fundamental idea that customers exposed to an ad might buy differently in response to an ad based on their individual characteristics. In fact, research shows, many customers seeing a non-tailored ad will buy less than those not seeing an ad.
Personalized marketing can uncover these differences and substantially increase the return on marketing investments in many settings such as retail and ecommerce, services marketing, business-to-business marketing and brand management. Healthcare companies should consider dedicating more resources to machine learning, which can power data-driven patient-centric outreach programs. Because individual health is a civic good, policy makers and organizations need to support these personalized outreach programs.
As for patients themselves, giving detailed personal data to a doctor or receiving highly personalized, unsolicited phone calls legitimately can seem like an invasion of privacy. But Mittal's research shows, it measurably has the potential to save your life.