Ambient intelligence-based framework for quantitative evaluation of cell culture operations to support standardization and training

Introduction The widespread implementation of cell therapy has increased the demand for stable and reproducible cell manufacturing methods. However, many cell manufacturing processes rely heavily on manual cell culture operations, and operator training remains largely dependent on subjective assessment and tacit knowledge. This limits the establishment of training targets and hinders operational standardization. Recent advances in ambient intelligence have created new opportunities for digitizing human work processes and supporting their quantitative evaluation. In this study, we developed an ambient intelligence-based framework for digitizing, visualizing, and quantitatively comparing manual cell culture operations. Methods A defined mock seeding operation was performed by one expert with 10 years of cell culture experience and three trainees with approximately two years of experience. Videos of the operation were recorded inside a biosafety cabinet, and the dominant-hand motion was extracted using an artificial intelligence (AI)-based skeleton estimation. The resulting time-series data were converted into quantitative speed-based temporal profiles for each operation, termed Operation Profiles. To compare operations with different execution times, the similarity between each Operation Profile and the target Operation Profile derived from the expert was evaluated using fast derivative dynamic time warping (fastDDTW). Distance and gap index were used as complementary metrics to assess the waveform deviation and temporal misalignment, respectively. Results Skeleton estimation AI, combined with visual quality control, interpolation, and manual correction, extracted the dominant-hand motion from the operation videos and enabled the construction of interpretable Operation Profiles that reflect the internal temporal structure of the defined cell culture operation. Although all operators performed the same operation, their Operation Profiles differed, indicating operator-dependent variations in the operation patterns. Iterative feedback based on phase-specific differences from the target expert operation was associated with progressive improvement in the trainees’ similarity to the target Operation Profile. Conclusions This study established a framework for recording, digitizing, and quantitatively evaluating cell culture operations using ambient intelligence. These findings demonstrate that the alignment of cell culture operations toward a target expert operation can be quantitatively evaluated through an Operation Profile analysis. By enabling objective comparison of operator behavior and quantitative evaluation of alignment toward a facility-defined target operation, this approach may provide a basis for operational standardization and training support in cell manufacturing.