Rethinking Shewhart Control Charts for High-Frequency Automated Data in Biopharmaceutical Manufacturing

Naveenganesh Muralidharan, Founder & Principal Consultant, Bench2Batch CMC Lifecycle Partners™

Abstract

High-frequency automated process data in modern biopharmaceutical manufacturing provide unparalleled visibility but also present new challenges for process monitoring. Traditional Shewhart control charts rely on the assumption that data are statistically independent and identically distributed. However, in automated environments where measurements occur at sub-minute intervals, consecutive readings are strongly autocorrelated. This violates Shewhart assumptions, resulting in artificially narrow control limits, inflated false alarm rates, and the risk of masking real process shifts. This article examines the underlying statistical issue using practical analogies, summarizes published evidence on the impact of autocorrelation, highlights relevant regulatory flexibility, and proposes the use of simple statistical summaries—such as minimum, maximum, and average values as a fit-for-purpose alternative monitoring strategy.

Introduction

The use of Shewhart control charts is deeply embedded in pharmaceutical manufacturing and quality assurance as part of statistical process control and Continued Process Verification (CPV) programs. However, many modern biopharmaceutical facilities have evolved toward automated operations, with process control systems capturing data at sub-minute intervals for critical parameters such as temperature, pH, dissolved oxygen, and agitation.

While this high-frequency data provides exceptional visibility into process behavior, applying traditional Shewhart charts directly to these datasets can produce misleading results. Specifically, the underlying statistical assumptions of Shewhart charts are violated, which leads to distorted control limits, unnecessary investigations, and potential misinterpretation of true process performance.

The central issue is autocorrelation the relationship between successive measurements in a time series. This article provides a clear explanation of autocorrelation using an intuitive temperature probe analogy, summarizes evidence on its effects on control charts, discusses regulatory perspectives, and outlines a simplified monitoring alternative based on basic statistical summaries.

Understanding Autocorrelation: A Practical Analogy

To understand autocorrelation in this context, consider monitoring the temperature of a bioreactor every five seconds. The temperature will not fluctuate abruptly between measurements; instead, each new reading will be very similar to the previous one. This is because the physical system and its control loops are designed to maintain stability and respond gradually. The degree of similarity between consecutive readings can be quantified by the lag-1 autocorrelation coefficient, a number between 0 and 1. A value near 1 indicates that each new data point closely follows the previous one – a behavior typical in automated control environments. In biopharmaceutical processes, lag-1 autocorrelation coefficients commonly range from 0.8 to 0.99.¹

This strong dependency between successive data points means that short-term differences in the dataset are artificially small. When these differences are used to calculate control limits in Shewhart charts, the limits become unrealistically narrow. As a result, normal fluctuations are flagged as “out-of-control”, even when the process is stable.²

Impact of Autocorrelation on Control Chart Performance

The effects of autocorrelation on control charts have been well documented in the statistical quality control literature. Kovářík and Briš investigated Shewhart chart behavior using simulated autoregressive processes and real industrial datasets.³ They demonstrated that:

• Even moderate autocorrelation significantly increases the false alarm rate relative to the nominal design.
• Control charts applied directly to autocorrelated data underestimate true process variability, biasing process capability indices downward.
• In a chemical process example, traditional Shewhart charts produced multiple “out-of-control” signals that did not correspond to actual process disturbances.

These findings highlight the mismatch between classical control chart theory and modern automated process data. When applied without modification, Shewhart charts can overwhelm operations teams with nuisance alarms, diverting attention from meaningful events.

Why Frequent Sampling Amplifies the Problem

Frequent sampling exacerbates the impact of autocorrelation. For example, consider a temperature control loop set to 37 °C. Over the course of an hour, natural process variation may be ±0.5 °C. However, if measurements are taken every 5 seconds, the differences between consecutive readings may be as small as ±0.02 °C. If control limits are calculated using these compressed short-term differences, the control chart becomes hypersensitive, signaling alarms for normal, stable behavior. This does not indicate poor process control; rather, it reflects the inappropriate application of a statistical tool designed for independent data to a dataset that exhibits strong serial dependence.2

Regulatory Flexibility and Fit-for-Purpose Methods

Importantly, regulatory guidance does not mandate the use of Shewhart control charts and allows manufacturers to adopt statistically sound alternatives that are appropriate for their data characteristics for CPV or process monitoring. The US Food and Drug Administration (FDA) encourages manufacturers to use statistically sound and risk-based approaches appropriate to their data and processes. The FDA’s Guidance for Industry: Process Validation emphasizes that statistical methods should be selected to suit the characteristics of the process under evaluation.⁴

Similarly, Tébar, Sáez, and O’Connor, writing on CPV modernization in the context of Pharma 4.0, emphasize that as manufacturing transitions to automated, high-frequency data environments, monitoring tools must evolve accordingly. Traditional SPC methods should not be applied uncritically to data for which they were not designed.⁵

This regulatory flexibility allows manufacturers to adopt alternative methods provided they are scientifically justified and documented – to better represent true process behavior.

Practical Alternative: Statistical Summary Monitoring

For continuously controlled parameters with high-frequency measurements, a simplified statistical summary approach can be more appropriate than applying Shewhart charts directly.

Instead of evaluating every data point individually, manufacturers can summarize data over meaningful operational periods (e.g., batch phases, hourly segments, or shifts) using basic statistics such as minimum, maximum, and average values.⁶

• Minimum and maximum values provide information about the operational envelope of the parameter over time.
• Average values indicate central tendency and allow monitoring for sustained drifts.
• Thresholds for these statistics can be based on engineering tolerances, historical performance, and risk assessments, rather than compressed short-term variability.

By focusing on summary statistics, nuisance alarms can be greatly reduced while maintaining sensitivity to real deviations. This method is straightforward, requires minimal computational complexity, and is easy to explain to both operators and regulators.

Conclusion

High-frequency automated process data are fundamentally different from the type of independent, low-frequency sample data for which Shewhart charts were originally designed. The presence of strong autocorrelation between successive data points violates the independence assumption, leading to narrow control limits and frequent false alarms.

Regulatory frameworks support the use of fit-for-purpose methods, allowing manufacturers to adopt monitoring strategies that accurately reflect process dynamics. Using simple statistical summaries such as minimum, maximum, and average values over relevant operational periods offers a practical, transparent, and defensible alternative to traditional Shewhart charts for CPV and routine process monitoring of automated data streams.

Subscribe to our e-Newsletters
Stay up to date with the latest news, articles, and events. Plus, get special
offers from American Pharmaceutical Review delivered to your inbox!
Sign up now!