Life sciences: Increasing speed-to-insight in pharma
AI and innovation in analytics
The use of analytics in evaluating pharmaceuticals is well-established and will be extended through the use of AI, according to Mark Lambrecht, global director of health and life sciences at SAS. “AI will be used operationally to increase efficiency and automation in clinical trials,” Lambrecht noted, “and in actions such as evaluating medical images in a few seconds rather than 20 minutes.”
Lambrecht highlighted several misunderstandings about analytics and AI that may present barriers. One is that a lot of data is needed. “You need to apply the right methods to the right problems,” he pointed out, “but you don’t necessarily need very large datasets to get meaningful results.” The other is that AI alone can give an exact result. “It is still very much a statistics game,” he remarked. “You can say that one patient has a higher risk than another, but exact predictions are not yet feasible.” As AI moves into full production, matching expectations to what can be delivered will be an important component.
Innovation is taking place in the clinical trial process through other means as well. “Virtual or remote trials are taking place without central monitoring,” Lambrecht said. “For example, patients entered a trial based on their use of an Apple watch.” In one study coordinated by Stanford University, more than 400,000 subjects wearing Apple watches used its heart sensor to detect irregular heartbeats, an indicator of atrial fibrillation (AFib). Among the participants receiving warnings, 84% were found to have had an AFib event, and upon further testing, one-third of those were found to have the condition.
Adverse event analysis
Information associated with adverse events can be particularly challenging to process. Much of it is in text form, with data being submitted through a variety of channels, including fax, email, phone calls to contact centers, and online forms. It arrives in different formats, including emails, spreadsheets, PDFs, and MS Word and text documents, as well as scanned documents.
One large global pharmaceutical company was struggling to process tens of thousands of these reports manually within its 7-day reporting requirement and assign them to the proper coding hierarchy, which contains 77,000 categories. The variability in the volume of reports posed a staffing problem. The requirement to monitor each case over time, once it was identified, further complicated the process.
In order to automate the reporting of adverse events, the pharmaceutical company chose PolyAnalyst from Megaputer, a text analysis solution. PolyAnalyst extracts primary information such as date and type of adverse event. Aiding in the semantic analysis are medical and pharmaceutical ontologies and a dictionary with classification terms. The system finds the appropriate MedDRA (an international standard for medical terminology) codes to classify adverse events for reporting purposes.
“To correctly detect and classify adverse events, the system has to resolve several layers of complexities,” said Sergei Ananyan, CEO of Megaputer. “First, it needs to identify patterns that might signal an adverse event, which frequently requires converting vernacular language to specific medical terminology present in MedDRA ontology. Second, it needs to make sure that the discovered issue is not being negated or just checked, and does not represent the indication for prescribing the drug or a fact from the past medical history of the patient.” Finally, PolyAnalyst maps the detected specific terms from lowest level of MedDRA ontology to the corresponding preferred term for classification, determines whether the event is serious or non-serious, and sets up a case for ongoing monitoring.
The system is trained on a large collection of adverse event reports, with the training results being validated by a human. “PolyAnalyst combines NLP and semantic rule-based techniques with innovative techniques based on artificial intelligence,” noted Ananyan. “The joint application of these two very different approaches helps the system achieve sufficiently high recall and precision of the results simultaneously.”