Arup Das co-wrote this blog post with Shankar Viswanathan.
Looking back at 2017, artificial intelligence and machine learning made impressive progress when it comes to improving cancer diagnosis and treatment. Using deep learning, computers are scouring images to detect signs of breast cancer in mammograms earlier than humans are currently capable of. Using AI “random forests,” a learning algorithm, investigators are more accurately predicting which drug combination will work better in BRAF mutant melanoma. The AI-facilitated discovery of Berg Health’s BPM 31510, a pancreatic cancer drug, has entered human development clinical trials. AI is not only powering oncology drug discovery, faster detection and personalized treatment but also helping to improve oncology commercialization effectiveness and agility through analytics.
Not only is oncology data getting better in quality (coverage, granularity, connectivity and lag), but also the rise in data collection sensors, natural language processing (NLP) and augmented reality is creating new, important data for the oncology companies to analyze. The new breed of learning algorithms powering this work can provide powerful advantages today. Lots of oncology sales and marketing issue areas have started to incorporate AI-powered algorithms, like supervised learning, to help solve patient, physician and provider analytics issues, and to predict stakeholder behavior. Here are three ways in which AI is affecting oncology analytics:
1. Physician analytics: As the market gets crowded with drugs with novel mechanisms of action, and precision medicine becomes the norm in oncology treatment, it’s very important for oncology companies to determine the key influencers and the social circle they influence so that messaging and sales force effort toward brand promotion can be optimized further. Launches will be more successful if these influencers are identified up front. It can shape everything from the engagement strategy to the segmented messaging of their networks.
In oncology, typically the doctors practicing at top oncology academic centers or centers of excellence are at the forefront of scientific exploration and the adoption of novel therapies. However, it’s usually difficult to determine the social circle of oncologists whom they end up influencing. Inferences using referral data/diagnosis are weak because of poor data connectivity across oncologists. Many of these professionals and other attendees are active on social media platforms like Twitter and Doximity, and in online cancer discussion forums. Using the active social media data, NLP techniques can identify the social circles and primary influencers among prescribers on various topics. Further discovery of social circles by creating a network graph of engagement between the social media participants among prescribers and running community detection algorithms is very feasible. This can lead to the identification of top influencers among prescribers by running a ranking algorithm on network graphs.
2. Provider analytics: How can AI help preempt protocol change for big providers? Most of the large cancer treatment centers and networks have standard protocols for treating patients in order to optimize the delivery of care and improve patient outcomes. Protocol standardization is becoming more common as most of the drugs are getting priced north of $125,000 and are offering less differentiation within a class of mechanism of action. For drugs that are clinically close in terms of efficacy and cost, these centers may choose one drug over the other based on real-world evidence. If outcomes (response rate, duration of therapy, side effects, etc.) are declining for a subset of patients at these centers, oncology companies can take preemptive action by having a dialogue with the centers before the center deprioritizes the use of the drugs altogether. AI can help scour the RWE data and call out the drugs that will have poorer outcomes based on historical drivers.
Typically, these large oncology institutions (consisting of integrated delivery networks and large corporate parents), contribute to approximately 30 to 50% of branded drug revenue in the U.S. Not only are these centers at the forefront of using novel therapies, but also they influence the drug usage in the smaller oncology communities around them. If oncology companies can flag this protocol change due to poor predictive patient outcomes in these centers, it will surely help them to intervene on time and protect the brand shares.
3. Patient analytics: Patient drop-offs account for a significant portion of lost revenue for brands given the significant cost of acquisition of new patients in oncology, where most of the newer therapies are getting approved in niche patient segments with specific genomic requirements. Incorporating the longitudinal sequences in the patient’s treatment history recurrent neural network can help identify the sequences in a patient’s treatment history. This will enable accurate prediction of the patient’s drop-off event. Hence, corrective measures can be taken to prevent the patient from dropping off, and interventions can be done to counter the drivers for the drop.
Although commercial teams can’t intervene and course correct in all drop-off situations, drop drivers related to improper infusion administration, sub-optimal side effect management, and access and reimbursement issues can definitely be taken up by the commercial teams, and the drop rates can be stemmed if these drivers and cases are known ahead of time. For example, if issues are around side effect management, timely education and support programs can be instituted up front. Obviously, a longer duration of therapy has direct implications on lifetime value and maximizing efficacy potential.
Especially in tumor types for which the data is thin due to low incidence, predicting therapy change can help optimize sales force cost and make targeting more focused. So how can AI help? Using longitudinal patient claims transactions, pharma companies can construct patient journeys. Aggregating patient journeys across multiple patients prepares the data set to discover significant co-occurrence patterns for different regimens, tests, diagnoses, treatments, adverse events and the length of duration. Leveraging non-linear, temporal-memory-based models (deep learning), companies can derive patient representations. Patient representations are a unique, fixed-length, real-value stream of a number of patient journeys constructed based on event co-occurrence patterns. Further, it uses supervised learning algorithms to predict time-to-line change. Similar constructs can encode patient journeys to predict patient progression and time-to-line change.
Timely prediction of line change generates alerts to reps and suggests appropriate actions for each case. For launches in smaller tumor types with low peak sales, it isn’t optimal to have a standing team of sales representatives as the patients are sparse. Hence, an alert-based strategy providing real-time inputs to reps can improve promotional effectiveness. A learning and adaptive setup for rep feedback can be used to improve the quality of alerts, and the system can send alerts to the field with sufficient lead time to make an impact before a treatment decision is made. Hence, reps can be in front of the oncologist at the right time with the right message.
Using AI, oncology companies have the power to target physicians dynamically just before they see patients so that the rep visits at just the right time with the right content. A combination of data-scouring using machine learning and supervised algorithms has the potential to intervene before the patient drops off or the centers make treatment protocol changes based on RWE data. With an understanding of drivers for drop-offs, careful planning and the use of powerful AI tools, oncology companies can eliminate anxiety and uncertainty, and find success with high ROI from their sales and marketing efforts. AI tools essentially extend the opportunities for the oncology companies and empower them to respond rapidly.