Machine learning recovers corrupted pharmaceutical 3D printing formulation data

Pharmaceutical 3D printing is an emerging digital manufacturing technology capable of autonomously producing personalised medicines. However, the same reliance on digital workflows that enables this innovation also introduces new vulnerabilities, most notably the risk of cyberattacks. In such scenarios, malicious actors could corrupt data relating to the medicine to be printed, either by deleting or modifying critical information or introducing subtle noise that is difficult to detect; either could potentially compromise patient safety. To begin to address this challenge, here we investigate the application of machine learning, specifically denoising autoencoders (DAEs), for the reconstruction of corrupted pharmaceutical formulation data. The dataset comprised 1,623 formulations (with 336 ingredients), totalling over 545,000 individual data points. To simulate potential cyberattack scenarios, the dataset was corrupted deliberately in two ways: (1) random removal of 1 % - 50 % of the data points, mimicking targeted data deletion, and (2) introducing noise across all data points, simulating tampering or injection attacks. Multiple DAE configurations were evaluated to determine the ability to recover corrupted data. Performance revealed R² scores ranged from 0.989 ± 0.0017 at 1 % corrupted rate to 0.924 ± 0.0034 at 50 % corrupted rate. Further analysis revealed that DAE was accurately reconstructing both active pharmaceutical ingredient and excipient values, indicating that the model captured meaningful formulation relationships rather than memorising data patterns. Among the parameters tested, the learning rate was found to have a significant effect on DAE performance. In contrast, a traditional machine learning techniques failed to produce positive R² values across all data corrupted levels, further demonstrating the superior performance of the DAE. Therefore, DAEs have been shown to safeguard formulation data against data corruption. The work underlines the broader role of machine learning in enhancing digital resilience and maintaining data quality across the pharmaceutical sector.