Saama’s new auto-coding module improves medical coding accuracy using popular dictionaries.
Despite the importance of medical coding during clinical trials, accuracy rates have traditionally been very low: in the 30–50% range for adverse events (AE) and 50–60% for medications. The fact that unclassified items must go through another time-consuming manual query process for resolution just adds insult to injury.
While working with a top Pharma to help accelerate its COVID-19 vaccine clinical trial, Saama developed an innovative medical coding module that reduces manual workload and improves accuracy.
Part of Saama’s Smart Data Quality (SDQ) solution, the auto-coding module uses natural language processing (NLP) capabilities to code AEs and medications, and can even auto-generate queries for items that can’t be properly coded.
Saama recently tested the auto-coding module against current coding processes at another pharma company, which was only achieving 50% coding accuracy with MedDRA and 30% accuracy with WHODrug.
Through a combination of ground truth data, user feedback, and model training, Saama’s auto-coding module improved MedDRA accuracy by 150% and WHODrug accuracy by 250%.
| MedDRA Autocode | ||
| Current | Target | Saama |
| 50% | 70% | 88%* |
| WHODrug Autocode | ||
| Current | Target | Saama |
| 30% | 50% | 81%* |
*After feedback and additional teach cycles to train the machine learning algorithm, accuracy will keep improving over time.
How the Auto-Coding Module Works
Omitted Terms from an EDC application or external coding system (e.g., TMS), flow into the module for auto-coding or querying.
After deep learning models predict coding decisions for each term, a built-in workflow system assigns coding decisions to human users for approval or rejection and tracks the activities of authorized coders and their managers.
If a coding decision is rejected, the auto-coding module will log the feedback and the rejected terms will await additional human action, using a built-in browser that allows users to look through the selected dictionary.
If approved, the auto-coded terms flow back to the module via import APIs or flat file downloads, so a single source originates all final coding decisions. Terms generated from automatic queries also flow into the coding application.
Getting Started with Auto-Coding
As long as a company has access to one or more of the official medical dictionaries, it’s fairly easy to start using Saama’s auto-coding module in about four to six weeks.
In order to train the AI model and set it up effectively, pharma companies must be able to provide Saama with the following:
- ~20 historical study data sets with synonym lists
- At least one individual with knowledge of internal coding processes
- At least one or two people who can be trained to interact with the auto-coding module
- Technical support for access to and integration with existing coding tools
In addition to saving incredible amounts of time, SDQ’s auto-coding module can save pharma companies more than $200K per study/year. To learn more, contact us today.
