Quality risk management for microbial control in membrane-based water for injection production using fuzzy-failure mode and effects analysis

The cause-effect fishbone diagram

Identifying potential failure modes is a fundamental step in FMEA (Xu et al., 2020). The fishbone diagram, also known as the Ishikawa or cause-and-effect diagram, is a widely used tool for root cause analysis (Ito et al., 2022). Originally developed by Ishikawa in 1990, this quality control tool was designed to identify factors contributing to an overall effect and mitigate product quality defects (Coccia, 2020). In the present study, the fishbone diagram was employed to identify and analyze the root causes of microbial issues in membrane-based WFI systems.

Conventional FMEA

FMEA is a structured approach used to detect and mitigate potential failures or defects within processes, products, or systems (Sharma & Srivastava, 2018). Originally developed in the 1940s for the aerospace and military sectors to enhance system reliability and reduce the chances of failure (Demirkaya, 2022), FMEA has since been adopted across diverse industries, including automotive (Petrescu, Cazacu & Petrescu, 2019), healthcare (Abbassi, Brahim & Ouahchi, 2023), and manufacturing (Demirkaya, 2022), to optimize product and process design, improve reliability, and lower failure-related costs.

After identifying potential failure modes in a system, each was prioritized based on its risk priority number (RPN), which calculated by multiplying three risk parameters: severity (S), occurrence (O), and detection (D). Once assigned, these ratings yielded an RPN value through Eq. (1), where higher values indicated a higher priority for addressing the specific failure mode:

(1) $$\mathrm{R}\mathrm{P}\mathrm{N}=\mathrm{O}\times \mathrm{S}\times \mathrm{D}$$

The rating scale (Tables 1–4) assigns scores for severity, occurrence, and detection for each failure mode (FM) based on pharmaceutical experts’ insights.

While FMEA is a valuable tool for risk assessment and mitigation, it has certain limitations (Dai et al., 2011; Selvan et al., 2013; Septiyana, 2021; Sharma, Kumar & Kumar, 2005):

• i)

  FMEA assigns equal weight to severity, occurrence, and detection, which may not accurately reflect their relative importance in specific situations. For instance, a failure mode with high severity while low occurrence may be more critical than one with low severity and high occurrence.

• ii)

  Certain combinations of severity, occurrence, and detection scores can result in identical RPN values, even when the underlying risk profiles differ significantly.

• iii)

  The RPN values are limited by the scale ranges for severity, occurrence, and detection, leading to clustering of values (e.g., RPN 120 appears 24 times, while values, such as 1, 123, and 1,000 appear only once).

• iv)

  FMEA calculates RPN for each failure mode independently, overlooking any interactions or dependencies between failure modes.

Fuzzy-FMEA approach

To address the limitations of RPN values from conventional FMEA, this study applied fuzzy logic to FMEA, generating a fuzzy-RPN (fRPN) that overcomes these limitations and enhances the interpretability of the findings.

The Fuzzy-FMEA approach enhances traditional FMEA by incorporating fuzzy logic, which helps address common limitations found in conventional FMEA, including equal weighting of severity, occurrence, and detection parameters and the lack of accounting for dependencies among failure modes. By assigning flexible, linguistic values to risk parameters, the Fuzzy-FMEA provides a more nuanced assessment that can reduce the clustering of RPNs and avoid redundancy. Additionally, using fuzzy sets accommodates the subjectivity and variability in expert opinions, which often arise in risk evaluations in pharmaceutical water systems. However, the Fuzzy-FMEA requires a high level of expertise to define appropriate fuzzy sets and membership functions, which can make it resource-intensive compared to traditional FMEA. Moreover, the method relies on fuzzy rule bases, which may limit its accuracy if not calibrated correctly through extensive expert input.

Fuzzy-FMEA approach

A rule-based expert system approach is widely applied in fuzzy FMEA to address the limitations of conventional FMEA (Akyuz, Akgun & Celik, 2016; Goksu & Arslan, 2023). The foundation of this approach is a fuzzy rule base, a collection of IF-THEN rules that represent expert knowledge. In contrast to traditional methods relying on precise numerical values, fuzzy logic uses linguistic variables for input and output description. In the Fuzzy-FMEA applied in this study, severity (S), occurrence (O), and detection (D) serve as linguistic variables, with fuzzy sets defined as “remote,” “low,” “medium,” “high,” and “very high”.

This study employed the Mamdani inference method to integrate the three risk-indexed parameters (S, O, D) nonlinearly through fuzzy IF-THEN rules, generating a more notable RPN. The Mamdani inference process involves five main steps: constructing the rule base, fuzzification, rule evaluation, rule aggregation, and defuzzification. The Fuzzy-FMEA approach for this study is outlined in Fig. 1 (Ceylan, 2023; Goksu & Arslan, 2023).

i. Rule Base Construction

The first step of a rule-based expert system is to collect fuzzy IF-THEN rules from domain experts. Fuzzy rules are statements that relate input variables to output variables in a linguistic form, typically expressed as “IF… THEN…” statements. A typical IF-THEN rule in the Fuzzy-FMEA can be expressed as follows (Akyuz, Akgun & Celik, 2016; Goksu & Arslan, 2023):

(4) $$R_i:\mathrm{I}\mathrm{f}o\mathrm{i}\mathrm{s}{O}_i,s\mathrm{i}\mathrm{s}{S}_i,\mathrm{a}\mathrm{n}\mathrm{d}d\mathrm{i}\mathrm{s}{D}_i,\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{n},\mathrm{R}\mathrm{P}\mathrm{N}\mathrm{i}\mathrm{s}{R}_{i}i=1,2,...,K$$where R_i refers to the rule number, K is the total number of rules, O_i, S_i, D_i, and R_i are the fuzzy sets in O, S, and D, respectively; o, s, d, and RPN are the input and output variables of the rule base expert system, respectively.

ii. Fuzzification

Before starting the fuzzy inference process, the crisp inputs need to be fuzzified using the membership functions to obtain an activation weight for each rule (Akyuz, Akgun & Celik, 2016). Thus, the second step is to transform input values into membership degree quantities, mainly ranging from 0 to 1. This process is called fuzzification.

iii. Rule Evaluation

Evaluating fuzzy rules is a critical step in the fuzzy inference process, where the activation weights for each rule are determined by assessing the membership degree of the input variables in the rule’s antecedents. This step involves identifying the relevance or applicability of each rule based on the current input values. The Mamdani method applies the minimum operator as fuzzy implication for this evaluation. The output fuzzy set R_i′ for each rule is by Eq. (5) (Akyuz, Akgun & Celik, 2016):

(5) $$\mathrm{\mu }{{\mathrm{R}}_{\mathrm{i}}}^{\prime }\left(\mathrm{R}\mathrm{P}\mathrm{N}\right)={\mathrm{\alpha }}_{\mathrm{i}}\wedge \mathrm{\mu }{\mathrm{R}}_{\mathrm{i}}\left(\mathrm{R}\mathrm{P}\mathrm{N}\right)1\le i\le \mathrm{K}$$where α_i = µ_Si(s) $\wedge$ µ_Oi(o) $\wedge$ µ_Di(d), representing the activation weight of each rule’s antecedent.

iv. Rule Aggregation

Once the weighted contributions of each activated rule are calculated, the next step is to combine these contributions to produce the final fuzzy output. This aggregation is achieved using the maximum operator as shown in Eq. (6):

(6) $$\mathrm{\mu }{{\mathrm{R}}_{\mathrm{i}}}^{\prime }\left(\mathrm{R}\mathrm{P}\mathrm{N}\right)=\vee \mathrm{\mu }{{\mathrm{R}}_i}^{\prime }\left(\mathrm{R}\mathrm{P}\mathrm{N}\right)i=1,2,\dots ,\mathrm{K}$$v. Defuzzification

Defuzzification converts fuzzy outputs from the inference process into crisp values, a necessary step when prioritizing the outputs of a fuzzy system. Several methods exist for defuzzification, including the centroid, weighted average, and bisector methods. This study used the centroid method due to its widespread application in the literature. The corresponding formula for this method is denoted as Eq. (7):

(7) $${\mathrm{R}\mathrm{P}\mathrm{N}}^{″}={\displaystyle \frac{\sum _{j=1}^l\mathrm{\mu }{\mathrm{R}\mathrm{i}}^{\prime }\left(\mathrm{R}\mathrm{P}\mathrm{N}\mathrm{j}\right)\times \mathrm{R}\mathrm{P}\mathrm{N}\mathrm{j}}{\sum _{j=1}^l{\mathrm{\mu }\mathrm{R}\mathrm{i}}^{\prime }\left(\mathrm{R}\mathrm{P}\mathrm{N}\mathrm{j}\right)}}$$where RPN″ represents the estimated value, and l denotes the number of activated rules, RPN_j.

Problem description

As discussed previously, microbial control is a significant challenge for a qualified membrane-based WFI system, and each critical component poses potential risks for microbial breakthrough. This study proposed a proactive approach to mitigate malfunctions in a typical membrane-based WFI system, comprising a conventional purified water generation unit and UF in the distribution loop. This system frequently triggers alarms due to microbial breakthrough issues. The following sections outline the evaluation of these failure modes using the proposed method and present strategies for preventing major defects proactively.

Application of the proposed method

To address the main microbial hazards in membrane-based WFI systems, this study combined pharmaceutical industry expertise with relevant guidance documents. Through a cause-effect fishbone diagram and brainstorming, 24 hazards were identified across different system components. Figure 2 displays the cause-effect fishbone diagram, while Table 5 presents a detailed FMEA worksheet. This worksheet encapsulates expert insights and draws from the Good Practice Guide: Membrane-based Water for Injection Systems, concentrating on the potential microbial hazards inherent to these systems (ISPE, 2022).

Effective design and operation of membrane-based WFI systems require a thorough microbial risk assessment to identify contamination hazards. Preventive and control actions are necessary to prevent microbial proliferation and enhance system robustness. The FMEA worksheet provides a practical overview, listing identified hazards and analyzing the potential failure modes and causes of each. For instance, the RO system presents a major hazard, contributing to failure modes FM7, FM8, and FM9. For FM9, the effect is “reduced efficiency in removing inorganic, organic, and microbial impurities, potentially impacting the microbial and endotoxin levels in product water,” caused by “frequent RO system use or inadequate disinfection leading to clogging or poor feed water quality.

To evaluate the risk level of each failure mode, Fuzzy-FMEA was employed to calculate the RPN. Pharmaceutical experts rated each failure mode for severity, occurrence, and detection based on their experience and knowledge. The mean of these ratings was utilized to calculate the fRPN values in MATLAB, prioritizing them from the highest to the lowest.

In MATLAB’s fuzzy logic program for FMEA, three input variables, including severity, occurrence, and detection, are defined, along with one output variable, fRPN. This study involved three senior pharmaceutical water system practitioners to evaluate each failure mode, and the results are presented in Table 6. Severity and detection use five-level triangular MFs, while occurrence employs a five-level trapezoidal MF, and fRPN has a ten-level triangular MF (Figs. 3–6).

Each input variable has five descriptors (very low, low, medium, high, and very high), totaling 15 descriptions, while the output variable fRPN includes ten levels (none, very low, low, high low, low medium, medium, high medium, low high, high, and very high). The fuzzy rule base, with 125 unique rules, is presented in Appendix 1. For instance, Rule 16 indicates that with remote occurrence, high severity, and certain detection, the RPN is low. In contrast, Rule 50 indicates that when occurrence is low, severity is very high, and detection is very low, then the resulting RPN is rated as high.

Results and Discussion

The findings revealed notable differences in risk ranking between the two assessment methods. Traditional FMEA has limitations, which can lead to inefficiencies and increased costs due to excessive response measures. For example, FM11 and FM16 have received the same RPN scores in traditional FMEA were ranked differently when assessed through Fuzzy-FMEA, providing a more informed basis for preventive measures. To mitigate this, this study utilized fRPN values to develop more scientifically based and balanced preventive and control strategies, along with optimized recommendations.

Expert evaluations indicate that the fRPN values for failure modes in both the distribution and production systems are relatively high. This is due to the close correlation between the distribution system and the drug production system, where the failure modes of each influence the other significantly. Additionally, certain failure modes in the pretreatment system, such as FM3 and FM2, also exhibit high fRPN values, highlighting their remarkable impact. The Fuzzy-FMEA method provides several practical advantages for microbial risk assessment in WFI systems. By applying fuzzy logic to severity, occurrence, and detection ratings, it can capture subtle variations in expert judgment, leading to a more refined prioritization of failure modes. This method is particularly beneficial in complex systems, such as membrane-based WFI systems, where the potential interactions between failure modes and microbial hazards are multifaceted. However, despite its benefits, Fuzzy-FMEA has limitations. Firstly, the method requires sophisticated software tools and computational resources, and the accuracy of its results heavily depends on the quality of the fuzzy rules and membership functions defined by experts. Additionally, while fuzzy logic provides flexibility, it may introduce ambiguity in cases where membership functions overlap significantly. As a result, practical application of Fuzzy-FMEA necessitates thorough validation and recalibration based on real-world system feedback to ensure reliable risk prioritization.

Table 7 demonstrates that FM14 has the highest RPN, reaching 7.00. The ultrafiltration (UF) system, serving as a microbial barrier, is typically positioned at the end of the ambient distribution system before the point of use to ensure that product water meets microbial and endotoxin standards (ISPE, 2022). However, retained substances on the UF surface can act as nutrients for viable bacteria in the distribution system, potentially leading to biofilm formation on membrane and pipeline surfaces. Consequently, improper UF system design or operation may turn it into a contamination source in the distribution system, impacting the microbial and endotoxin quality of the product water. Furthermore, if the distribution system’s sterility cannot be assured, placing UF before the point of use may not suffice to meet microbial and endotoxin requirements.

FM16 is also a significant failure mode in the distribution system, with the second-highest RPN of 6.73. Any failure mode introducing external contamination can negatively affect the product water’s microbial quality, as the introduction of bacteria can lead to biofilm formation in the distribution pipeline (Collentro, 2016).

The RPN values rank FM7 as the third and FM9 as the fourth. The RO system, essential in the production process, removes ions, total organic carbon (TOC), bacteria, and endotoxins, thereby helping to prevent biofilm formation (ISPE, 2022). The O-ring, installed at the pressure vessel’s end-cap and in RO component connectors, prevents bypass contamination; however, insufficient sealing can compromise bacterial removal, causing bypass contamination (Fujioka et al., 2019; Liu et al., 2013; Pype et al., 2016). Additionally, membrane fouling significantly reduces RO membrane filtration performance, impacting both production flux and filtration efficiency (AlSawaftah et al., 2021).

FM15 ranks fourth among all the failure modes. The vent filter on the storage tank enables air or gases to flow in and out while acting as a microbial barrier to maintain sterility. Studies indicated that the filter material in respirators could undergo oxidation and degradation after 3–6 months of exposure to high temperatures and ozone, requiring regular filter replacement (Jornitz, Jornitz & Meltzer, 2008). If not replaced promptly, prolonged exposure could cause the organic materials in the filter to degrade, reducing its effectiveness against microorganisms (Tidswell & Bennett, 2017).

Another critical failure mode, FM11, involves continuous electrodeionization (CEDI), which is typically installed downstream of the RO system for polishing deionization. CEDI also supports microbial control due to its electric field and the ionization of water into OH⁻ and H⁺ ions (ISPE, 2022). Improper pressure balance, however, can lead to membrane or seal failure, resulting in either internal leaks between concentrate and dilute channels or significant external leaks (ISPE, 2022). Malfunctioning CEDI can fail to reduce TOC and inactivate microorganisms, raising the risk of biofilm formation in the distribution system, making proper CEDI operation essential.

Fouling on membrane surfaces can be managed by a well-designed pretreatment system, minimizing fouling risks (Abushawish et al., 2023). A key component in pretreatment is the carbon filter, which has a high fRPN ranking. The media’s large surface area facilitates microbial growth, and granular activated carbon (GAC) provides approximately 35,000 m²/m³ of surface area for bacterial attachment through pores larger than 500 nm in radius (Knezev, 2015).

Research on microbial activity in GAC has demonstrated that fine GAC particles in filtered water could harbor bacteria concentrations 2–3 orders of magnitude higher than in the filtered water alone (Alves et al., 2019; Kempisty et al., 2019; Lu et al., 2021; Oh et al., 2016; Tyagi et al., 2020). Additionally, bacteria on GAC surfaces are shielded from UV radiation and chlorine disinfection due to protective adsorption sites on the carbon, limiting the effectiveness of these disinfectants (Atabaki, Idris & Siong, 2014; Patel et al., 2020). Consequently, activated carbon filters are generally not recommended in membrane-based WFI systems.

Preventive actions and recommendations for optimization process

Given the critical importance of microbial limit compliance in WFI, it is crucial to implement preventive and control measures targeting high-risk failure modes in membrane-based WFI systems. Besides adopting preventive strategies, process optimization is equally important to reduce microbial risks and ensure compliance.

To minimize contamination, it is advisable to avoid filters with large filter surfaces, such as activated carbon filters and softeners, which are particularly susceptible to microbial buildup. Installing medium-pressure ultraviolet (UV) light before the CEDI unit can effectively inactivate bacteria, helping to prevent microbial growth in the CEDI itself.

System configuration should prioritize minimizing conditions that encourage microbial growth. This includes removing organic and inorganic matter that can act as nutrients during pre-treatment and production, limiting surfaces that facilitate microbial attachment, and ensuring nearly sterile water enters the distribution system. By implementing these measures, a membrane-based WFI system can reliably produce water meeting strict quality standards.

This study established preventive measures and process optimization strategies for membrane-based WFI systems based on pharmaceutical industry experts’ insights and extensive experience. These recommendations are summarized in Appendix 2.