Opportunities for Machine Learning application in Pharmaceutical Sector

Clinical Development One crucial activity in the R&D is the clinical development of a drug to study it and demonstrate its clinical benefit from the discovery stage to regulatory approval from Health Authorities. One potential solution to decrease the cost and the timeline of clinical development is speeding up the patient enrollments and patient inclusion success rate in clinical studies by applying machine learning (ML) in this process. The pharma company (sponsor) in association with the clinical site (hospitals, etc) can use patient’s medical records to gather information on the patient profile and identify potential study subject for participation in clinical studies. It is possible in those clinical sites where the patient’s medical records are saved in an integrated digital platform containing all their data (the feature vector), like age, gender, clinical exams results, clinical symptoms, etc. In this case the sponsor/clinical site builds a supervised ML including the training set that would be the key inclusion criteria of the clinical study (e.g adults with specific laboratorial results and specific symptoms/signals). Then, the system analyzes this medical platform to classify potential subjects and flags it to the clinical site which will be able to engage with the patient for the subsequent activities. The manual identification of potential subjects has lots of pitfalls, as diagnosis failure or subjectivity, and applying the AI the patient enrollment might be accelerated and more efficient. Currently, there are those kinds of digital platforms to manage patient’s medical reports and the success factor is having the minimum feature vectors necessary to classify the patient as a potential subject for the specific study. The minimum amount of patient included in this platform will vary depending on the disease (common vs rare disease) and the drug’s mechanism of action ranging from few hundreds to some thousands. Finally, considering that the clinical sites will provide service to a broad and big population, being frequently updated with their medical data, the system would have enough regularity in it. Product materials When the product is approved by the Health Authority and go to the market, the company must increase the awareness about its beneficial profile and differs new product from the existing treatments. Several materials in different formats and with different approaches are created to disseminate the product’s information. One of the challenges is to gain and keep the physicians’ attention to educate them on the new drug characteristics and make them comfortable to prescribe it. A personalization algorithm can tailor the educational materials for each HCP depending on their profile and browsing historical data. Pharmaceutical industries compile data from the physicians who its field force (sales rep, MSL and others) engage with. The database is built to save objective characteristics, as medical specialty and workplace, as well subjective information, such as specific comments and insights gathered during the interaction with the company representative. This dataset with thousand examples is used to train an unsupervised learning which will cluster the physicians. In addition, a supervised learning predicts the likelihood of a material being interesting to a physician by solving it as regression problem analyzing his/her historical browsing data. This preference is used as an example of a labeled profile and is applied to the group of similar profiles and indirectly labeling the other profiles in the same cluster. Regarding this semi-supervised regularity, the profile database and the browsing data are frequently updated capturing new patterns and recalculating the material preferences. Triage of questions in call center The Interactive Voice Response (IVR) is widely used in call centers across multi sectors. In the old phone system technology, the customer chooses one option in the menu by touching the phone’s tone keypad and has his/her call routed to the specific recorded message about that topic. The new versions of IVR are applying voice recognition to enable the inclusion of more menu options and, consequently, more automated responses. The company can use a supervised training using labeled audio files with the product names spoken with different ascents and pronunciation, then it groups these audio files and associates them to the specific response. The process of recording and labeling the audio files is not complex for single words and can be performed in house when there are few label options (e.g. products). For this reason, for the simple applications with single words 10-15 audio files per label is enough and each audio files contains a specific pattern of sounds that enables to recognize the word when the customer speaks it.