Chronovector

An algorithm built for Netflix. Applied to pharmaceutical prescribing trends. The algorithm wasn't new. The application was.

The problem with deciling

Pharmaceutical marketing in 2020 was data-rich and analytically conservative. HCP segmentation was dominated by a process known as deciling: rank doctors by prescription volume, group them into ten tiers, engage the top prescribers most heavily.

It answered one question: how much is this doctor prescribing right now? It answered nothing about where they were heading.

A doctor who moved fifty patients to a new therapy last quarter looks identical to one who has just started doing so. A high-volume prescriber who has been in slow decline for two years sits in the same decile as one who has been growing steadily. Deciling was a snapshot. It had no trajectory. An HCP starting to move patients to a competitor therapy was invisible in the data until the volume shift became large enough to drop them a tier, by which point the window for intervention had likely closed.

The source

The solution came from an unexpected direction. Netflix's machine learning recommendation algorithms are built to cluster viewing behaviors over time, not just current preferences. They group users based on behavioral trajectories, which is structurally the same problem we were trying to solve for HCP prescribing patterns.

In 2021, before the current wave of AI interest had fully arrived, we had been using machine learning for computationally intensive processing work. Once a model is trained, calculations that would take thousands of hours manually run in a fraction of the time. The question was whether a content recommendation clustering algorithm could be adapted for pharmaceutical prescribing trajectories.

We called the output Chronovector. The name describes what it does: clustering based on a temporal prescribing vector.

Training and validation

Funding required a client willing to invest in something that had never been done before. Takeda's Trintellix was experiencing quarterly declines driven by emerging competitors. The challenge was distinguishing doctors who had always prescribed lightly from those actively transitioning patients to alternative therapies. Chronovector could solve for this far more precisely than any equation-based approach. Trintellix was the right test case.

The model was trained not just on Trintellix prescription data but on a broader range of therapy prescribing patterns. More training data meant faster and more reliable cluster formation. Within four weeks, the model was producing sixteen clusters across four tiers, with aligned aggregated trending and confidence intervals for each HCP's cluster assignment.

Validation: 250 HCPs within the 95% confidence threshold were manually reviewed. 248 were correctly assigned.

The marketing application

Raw cluster outputs were distilled into three categories that could be acted on directly: Growing, Steady, Declining. These mapped to targeting groups across High, Medium, Low, and Base tier prescribers, each with a directional trend attached.

Sales reps were briefed on the specific trajectory of each HCP before conversations. Media and email teams could actively reach mid and low-tier prescribers showing early signs of migration to alternative therapies. Budget that had previously been spent maintaining high-volume growing prescribers was redirected toward those where ground was being lost.

Trintellix halted its quarterly declines and returned to growth.

The outcome

Chronovector became Klick Health's first machine learning product. It was subsequently sold to eight clients, generating $2.2 million in billings, with ongoing recurring revenue through cluster refreshes as prescribing patterns evolve.

The underlying algorithm was not new. Applying it to a domain where nobody had thought to use it was the work.

AI and machine learning deliver real impact when applied to specific business problems — not as a general capability, but as a precise answer to a question that couldn't previously be asked.