The advent of new generation machines, higher processing power and increased data, and usage of data analytics and artificial intelligence have seen increased uptake across the industrial segments. Data intensive industries such as Life Sciences, Supply Chain, and E-Commerce have been increasingly leaning towards AI, to make sense of data and process it to make meaningful and actionable.
Pharmacovigilance, is one such area of Life Sciences that deals with patient and medical safety data. AI- driven advances in drug development process have resulted in increased drug development and sales. In 2014, total pharmaceutical revenues worldwide had exceeded one trillion US dollars for the first time. It’s estimated that the global pharmaceuticals market will triple by 2060. With this overall increase in sales of drugs, reporting of side effects by various existing and new channels is also anticipated to increase.
Why is context of extracted entities important?
One of the most critical steps in Individual Case Safety Report (ICSR) would be to extract raw information from sources such as doctor’s prescription, laboratory reports, hospital records, and social media.
As can be seen, the diverse nature of sources leads to diverse vocabulary and dialects, which increases the complexity of verbatim. This complexity further increases when the information received contains verbatim with multiple events or mix of events, indications, procedures, investigation, and medical history. For example, in a medical narrative, there may be information or mention of multiple persons (such as doctors, healthcare professionals (HCPs), and patient’s relatives) besides the patient. Similarly, drug names may have a large diversity that not only includes multiple drugs but also the lexical diversity between generic name, brand name(s), and other naming conventions of the drug. These are potential processing challenges, where any inaccuracies can lead to dramatically different processing outcomes or decisions.
To understand a medical term, in addition to its context with respect to an event or medical history, multiple aspects need to be deciphered by the machine before processing. Table 1 below mentions few of these aspects and their examples, such as word forms, negation, synonym, and morphology.
|word Forms||pain, pained, pains|
1. There was no pain
2. Pain was not severe
How effective is the AI-based automation solution?
To address this diversity, an AI solution should be able to create a knowledge representation that gives entities, attributes, and their relationships. The system should also be able to learn inconsistencies and based on that create new weightage rules, allowing successful and accurate representation of information.
The AI solution needs to be intuitive having an oversight of complete processing. One such solution is implementing and indicating the confidence scores for derived and non-derived fields. This allows to keep track of extraction and inform the confidence score for extracted entities to business domain, which can be calculated based on parameters such as algorithm used for extraction, degrees of match with the domain ontology, and so on.
Field-level confidence scores help in fine-tuning the system largely. But, certain decisions such as straight through processing versus manual interventions, can be taken via case-level scores, which are based on aggregate field-level scores with weightage on fields depending on criticality, reportability and so on.
Further, the visual appearance of information can be made appealing on a pharmacovigilance system through traffic light color-coding (e.g. >90% confidence as green), to make it easy for the user to determine further action with a cursory look.
As an illustration, imagine a typical setup of AI-based pharmacovigilance, doing ICSR Processing. The system would be reading the medical documents and information details on side effects sent by HCPs or patients. The AI solution would be reading the source information, interpreting and creating a knowledge representation of the case. This will be done using the contextual rules, ontology, model representation etc. The AI solution could be configured based on human and data training such that any case with very high accuracy (eg. 99%), over a configured threshold, and similar to historically processed case, can go for straight thru processing, and come into Quality Review Check workflow.
Is AI the way to a robust future in pharmacovigilance?
We have discussed some of the challenges that are present and approaches that can be taken, while implementing AI in pharmacovigilance. In addition, the challenges pertaining to changing ontologies, regulations etc. have implications on the solutions. However, this in no way, limits the boundaries of the solution. In fact, putting in place an AI-enabled robust system that can handle such changes in a deterministic way, is the way forward.