3D projection and multivariate image analysis for quantitative visual modelling of mixability and rheological behavior of lactose powder.

Background: This study reports the use of multivariate time and image analysis of avalanche videographic data for quantitative visual modelling of mixability. Its usefulness, in mechanistically modelling a powder's rheological behavior in relation to mixing, was evaluated.

Methods: Particle size distribution (PSD) of a pharmaceutical grade lactose powder was modified to reflect commercially encountered variability. The PSD variants were rheologically distinct and had different mixability. Avalanche testing was performed on the modified lactose powders. Avalanche rheological properties (ARP) profiles and videos were collected for numerical and quantitative visual modelling, respectively. In quantitative visual modelling, videos captured were transformed into serial projected images. Important features of the projected images were extracted as eigen-images, to derive the avalanche rheological visual metric (ARVM). Mixability was modelled as a function of ARP or ARVM and the rotation speed.

Results: Relative to the ARP model, the ARVM models were highly interpretable. As a univariate expression of ARP, ARVM also possessed construct validity (r greater than 0.99, slope ≥ 0.96). Important rheological features of the lactose powders were holistically visualized within a single eigen-image which enabled the generation of simpler models (5 versus 34 variables for ARP model). The ARVM models predicted mixability of lactose powders with greater accuracy than the ARP model (relative root mean square error of external validation ≤ 3.30% versus 4.96%).

Conclusions: Quantitative visual modelling is a viable alternative to purely numerical approaches. Most significantly, the model's interpretability and concreteness enable manufacturers to readily understand the risk posed by PSD variability on manufacturing processes and swiftly take pre-emptive actions, without being mired in multivariate data complexity. In addition, the use of quantitative visual approach in time series imaging, for studying and monitoring industrial processes, could also be explored.