A decision analysis of cancer patients and the consumption of ready‐to‐eat salad

Abstract

Listeria monocytogenes is a foodborne pathogen of concern for cancer patients, who face higher morbidity and mortality rates than the general population. The neutropenic diet (ND), which excludes fresh produce, is often utilized to mitigate this risk; however, an analysis weighing the theoretical listeriosis risk reduction of produce exclusion aspects of the ND and possible negative tradeoffs has never been conducted. Consequently, this work constructed decision analytic models using disability‐adjusted life years (DALYs) to compare the impacts of the ND, such as increased neutropenic enterocolitis (NEC) likelihood, with three alternative dietary practices (safe food handling [SFH], surface blanching, and refrigeration only) across five age groups, for cancer patients who consume ready‐to‐eat salad. Less disruptive diets had fewer negative health impacts in all scenarios, with median alternative diet DALYs per person per chemotherapy cycle having lower values in terms of negative health outcomes (0.088–0.443) than the ND (0.619–3.102). DALYs were dominated by outcomes associated with NEC, which is more common in patients following the ND than in other diets. Switchover point analysis confirmed that, because of this discrepancy, there were no feasible values of other parameters that could justify the ND. Correspondingly, the sensitivity analysis indicated that NEC mortality rate and remaining life expectancy strongly affected DALYs, further illustrating the model's strong dependence on NEC outcomes. Given these findings, and the SFH's ease of implementation and high compliance rates, the SFH diet is recommended in place of the ND.

1. INTRODUCTION

Individuals undergoing cancer treatment are highly vulnerable to foodborne illness, particularly listeriosis. Treatment‐induced neutropenia (a state of decreased neutrophils) puts these individuals at higher risk of invasive infection with increased morbidity and mortality rates, compared to other immunocompromised groups (Goulet et al., 2008; Guevara et al., 2009; Mook et al., 2011; Silk et al., 2012). Invasive listeriosis can be severe, progressing into sepsis and central nervous system (CNS) infections, such as meningitis and encephalitis, with mortality rates up to 40% (Allerberger & Wagner, 2010; Choi et al., 2018; Farber & Peterkin, 1991; Goulet et al., 2012; Swaminathan & Gerner‐Smidt, 2007). Alarmingly, Listeria monocytogenes has been found and laboratory‐confirmed from fresh, or ready‐to‐eat (RTE) produce (Beuchat, 1996; Harris et al., 2003; Olaimat & Holley, 2012; Painter et al., 2013; Zhu et al., 2017), identified as the cause of several produce‐related outbreaks (Angelo et al., 2017; Centers for Disease Control and Prevention [CDC], 2016; CDC, 2021; CDC, 2022; McCollum et al., 2013; Stephan et al., 2015), and is attributable to over 70% of healthcare‐associated foodborne outbreaks (Boone et al., 2021), prompting the adoption of various food safety diets.

One such food safety diet is the neutropenic diet (ND), which eliminates the consumption of foods not pasteurized or cooked prior to consumption, including fresh produce and RTE salad (Heng et al., 2020; Sonbol et al., 2015). The ND is used as a risk management intervention in up to three‐quarters of hospitals (Braun et al., 2014; Carr & Halliday, 2015; French et al., 2001; Smith & Besser, 2000), despite having never been proven to reduce the risk of foodborne illness in cancer patients (DeMille et al., 2006; Heng et al., 2020; Lassiter & Schneider, 2015; Moody et al., 2018; Trifilio et al., 2012; van Tiel et al., 2007). Due to the dietary restrictions, patients report negative impacts on quality of life (McGrath, 2002; Moody, Meyer, et al., 2006) and low adherence rates (Heng et al., 2020; Lassiter & Schneider, 2015; Moody et al., 2018). Compounding this, Maia et al. (2018) demonstrated that the ND offers significantly less fiber and vitamin C than a regular diet and does not meet recommended values for fiber, vitamin C, calcium, or iron.

The effects of dietary fiber and vitamin C on intestinal integrity and colitis have been extensively described in the literature. Al‐Qadami et al. (2022) reported that gut bacterial fermentation of fiber is needed to generate short‐chain fatty acids, which alleviate gastrointestinal inflammatory response and reduce intestinal mucosa permeability. This was exhibited in mice models, where mice with mucositis that were given low and high fiber diets experienced increased and decreased intestinal permeability, respectively (Gallotti et al., 2021), and chronic and intermittent dietary fiber deficiency caused colonic mucosa deterioration and subsequent colitis (Desai et al., 2016). Likewise, proper doses of vitamin C were found to halt intestinal mucosa erosion and stimulate barrier recovery in guinea pigs (Qiu et al., 2021), and assuage enterocolitis symptoms, intestinal cell death, and inflammatory response in mice (Mousavi et al., 2020). As such, produce exclusion on the ND may contribute to gastrointestinal afflictions. This is supported by Gupta et al. (2022), who reported that neutropenic enterocolitis (NEC), a severe condition causing diarrhea and intestinal inflammation, is more common in patients on the ND than in patients on standard hospital diets, with morbidity rates of 33% for pediatric patients on the ND and 5% for those on a standard diet.

Not only are the benefits of the ND heretofore inconclusive, but its administration regarding RTE produce is highly varied between and within institutions, leading to unstandardized treatment across patient groups. First, complete elimination of fresh produce consumption is incongruent with guidelines set forward by the U.S. Food and Drug Administration (FDA) and American Cancer Society (ACS) (ACS, 2019b; FDA, 2020). Additionally, the number of hospitals that implement the ND may be anywhere between 29% and 77% (Braun et al., 2014; Carr & Halliday, 2015; French et al., 2001; Smith & Besser, 2000). Institutions also differ in when the diet is commenced, types of foods excluded, and discontinuation criteria (Braun et al., 2014; Carr & Halliday, 2015; French et al., 2001; Smith & Besser, 2000). However, a recent risk model for cancer patients who consume properly stored RTE salad demonstrated low listeriosis risk (on the order of 10⁻¹⁰ per chemotherapy cycle) associated with kitchen‐scale hygienic interventions, such as surface blanching (boiling for 25 s), or even rinsing, as advised by the FDA and ACS (Gomez et al., 2023).

It is currently unclear whether the potential benefits of the produce exclusion aspects of the ND (estimated listeriosis risk reduction) outweigh the aforementioned costs (increased rates of NEC, reduced intake of micronutrients, and lost quality of life). The same is true for less disruptive interventions, which may pose a marginally greater listeriosis risk, but a smaller NEC risk. The likelihood and magnitude of listeriosis health outcomes have never been contrasted to health outcomes associated with the ND's fresh produce restrictions. This study presents a decision analysis comparing risk–risk tradeoffs with other food safety interventions.

Decision trees can be used to graphically model difficult decisions. They present possible outcomes of the decision, accounting for accompanying probabilities, uncertainties, and consequences (Clemen & Reilly, 2013). The most desirable decision can be chosen by calculating the expected values (EVs) of each decision's outcomes and selecting the outcome with the highest or lowest EV. If the outcome is phrased in terms of negative value, as is the case in this model, then the most desirable decision will be the outcome with the lowest EV. When a decision model parameter is variable or uncertain, the “switchover point” can be calculated to ascertain the value of the parameter that leads to the other outcome having the lowest EV and therefore becoming the preferred intervention (Mitchell‐Blackwood et al., 2011). The switchover point is the level of a specific parameter that justifies the use of one intervention over another, when all other parameters are held constant. The objective of this study was to use decision analysis and the switchover point approach to compare risk–risk tradeoffs of the produce restriction aspect of the ND to three alternative intervention strategies applicable to RTE salad: (i) safe food handling (SFH), (ii) surface blanching, and (iii) refrigeration only, for five different age groups: (i) 0–4 years, (ii) 5–14 years, (iii) 15–44 years, (iv) 45–59 years, and (v) 60+ years.

2. METHODS

2.1. Exposure scenario and interventions

This study's exposure scenario considers cancer patients who consume RTE salad, as prepared by four food safety dietary practices. Patients with all types of cancer experience increased risk of invasive listeriosis infection. For this analysis, listeriosis risk estimates are conservatively estimated for hematological cancer patients, who have a higher listeriosis risk than other cancer patients (Gomez et al., 2023; Goulet et al., 2012; Mook et al., 2011); this means that the numbers provided by this model in terms of listeriosis risk are on the high end of the range of estimates for the average cancer patient. Listeriosis risk estimates and corresponding EV calculations were based on a previous risk model using the timeframe of one chemotherapy cycle and RTE salad consumption reported in g/day by cancer patients (Gomez et al., 2023; S. Ilic, personal communication, December 6, 2018). Four RTE salad food safety diets were considered: (i) ND (exclusion of all fresh produce), (ii) SFH (following FDA guidelines, notably refrigeration, rinsing, and drying), (iii) surface blanching (following FDA refrigeration guidelines and boiling non‐porous produce for 25 s prior to consumption), and (iv) refrigeration only (following FDA refrigeration but not rinsing guidelines). These interventions can all be applied at either the hospital or home scale, so it was assumed that they were followed throughout the entire chemotherapy cycle. For all interventions, it was assumed that salads consisted of RTE greens, cucumbers, and tomatoes refrigerated at 4°C. The amounts of L. monocytogenes persisting in salads following the various interventions were calculated in the previously noted risk model (Gomez et al., 2023).

2.2. Health outcomes

Following the decision to consume RTE salad, there are two potential health outcomes, (i) listeriosis and (ii) no listeriosis, with the possibility of NEC. Because every person making the decision in this study would have been diagnosed with cancer, they have the same risk in terms of cancer health outcomes, and the model appraises outcomes due to listeriosis and the chosen food safety dietary practice only. Based on previous listeriosis burden of disease studies (de Noordhout et al., 2014, 2017; Kemmeren et al., 2006), five non‐pregnancy‐associated listeriosis health outcomes were considered (pregnancy‐associated listeriosis has markedly different health outcomes than non‐pregnancy‐associated invasive listeriosis), (i) self‐limiting gastroenteritis (non‐hospitalized), (ii) curable gastroenteritis (hospitalized), (iii) CNS infection, (iv) long‐term CNS complications, and (v) sepsis (Figure 1), with potential death occurring as a result of sepsis or CNS infection only (Doorduyn et al., 2006; Goulet & Marchetti, 1996; Guevara et al., 2009; Kiss et al., 2006). Only systemic infections (CNS infection and sepsis) follow hospitalization, as gastroenteritis alone is considered noninvasive listeriosis, described as “mild” and/or “self‐limiting” with little to no mortality, only causing death when it progresses into more pervasive infections in immunocompromised individuals, such as cancer patients (de Noordhout et al., 2014, 2016; Ooi & Lorber, 2005; Seddon et al., 2018). The health outcome for no listeriosis (Figure 2) consisted of three states: NEC, death, and recovery.

FIGURE 1.

Listeriosis potential health outcomes.

FIGURE 2.

No listeriosis potential health outcomes.

2.3. Model metrics and formulas

The metric used to quantify health outcomes for each food safety diet was the disability‐adjusted life years (DALYs), an internationally recognized metric used since the 1990s to estimate global disease burden (Murray et al., 1996). DALYs account for years of life lost (YLL) and years lived with disability (YLD), representing mortality and morbidity, respectively, both of which are important to the model's targeted stakeholders, cancer patient caretakers. Within the YLD calculation (Equation 1), both duration and severity are considered, which makes the DALY appropriate for comparison of different disease risks. Similarly, the YLL calculation (Equation 2) can be adjusted to the remaining life expectancy (RLE) of the multiple target populations (i.e., accounting for age). It was prudent to use such a metric for comparison in this study because listeriosis outcomes are dominated by mortality (YLL), whereas NEC outcomes are driven by both morbidity (YLD) and mortality. Hence, use of the DALY ensures a comparable evaluation of this food safety diet decision. Conventional DALY formulas summing years lived with disability (YLDs) and YLL resulting from untimely death (YLLs) were used with no age weighting or time discounting (Equations (1), (2), (3)) (Devleesschauwer et al., 2014). Disability weights (Equation 1), used in YLD calculations, are measures representing the severity of a disease state, ranging from 0 (perfect health) to 1 (equivalent to death) (Global Burden of Disease Collaborative Network, 2020):

$$\mathrm{YLD}=\sum \mathrm{duration}\mathrm{of}\mathrm{ailment}{\left(\mathrm{years}\right)}_{\mathrm{i}}\times \mathrm{disability}\mathrm{weigh}{\mathrm{t}}_{\mathrm{i}},$$ (1)

$$\mathrm{YLL}=\mathrm{RLE}\mathrm{at}\mathrm{age}\mathrm{of}\mathrm{premature}\mathrm{death}\left(\mathrm{years}\right),$$ (2)

$$\mathrm{DALY}=\mathrm{YLD}+\mathrm{YLL}.$$ (3)

Time discounting (ascribing a lower value to future years of health lost) and age weighting (assigning lower weights to YLL for the young or old) were not utilized. These practices align with the latest World Health Organization Global Burden of Disease (GBD) Study methods, which state that the DALY is intended to quantify health outcomes as opposed to the social value of health outcomes (Global Burden of Disease Collaborative Network, 2020).

2.4. Decision model and analysis

Each decision model assumes that an individual is presently being treated for cancer. Decision trees were developed for 3 RTE salad intervention decisions, (i) ND versus no refrigeration only, (ii) ND versus SFH, and (iii) ND versus surface blanching, for 5 age groups, (i) 0–4, (ii) 5–14, (iii) 15–44, (iv) 45–59, and (v) 60+, as grouped in the 1990 GBD study (Murray et al., 1996), totaling 15 decision trees. Although children under two may not be eating RTE salads, it was important to consider the 0–4 age group in analyses as children may begin eating components of the model RTE salad product (greens, tomatoes, and cucumbers) as purée at under a year old, and to determine if this group necessitated a different food safety diet than the other groups.

Decision nodes, contrasting the ND with an alternative, were represented as square boxes (Figure 3). Chance nodes appear as circles and denote the probability of an event (outcome) occurring. For example, P ₁ and P ₁₀ represent the probabilities of acquiring listeriosis while on the ND and non‐ND interventions, respectively:

$$\mathrm{EV}\left(\mathrm{intervention}\right)=\sum \left(\mathrm{probability}\times \mathrm{associated}\mathrm{DALYs}\right).$$ (4)

FIGURE 3.

Decision tree for cancer patient consumption of fresh produce.

The other probabilities are assigned as follows: NEC while on the ND (P ₂), NEC while on the alternative food safety diet (P ₁₁), death following NEC (P ₉), listeriosis requiring hospitalization (P ₃), listeriosis with CNS infection (P ₄), listeriosis with sepsis (P ₅), death following CNS infection (P ₇), CNS infection long‐term complications (P ₈), and death following sepsis (P ₆). Triangles indicate branch terminals.

2.5. Model parameterization

Because no well‐developed point estimates existed for most of the model input parameters, and many are variable and/or uncertain, each parameter was represented by a distribution rather than a point estimate. As will be described later, a Monte Carlo simulation was used to run the calculations using these input distributions. A literature search was performed to parametrize the model. Only data concerning acquired (non‐perinatal) listeriosis cases were considered. Studies referring to sepsis as septicemia or bacteremia or CNS infection as meningitis, encephalitis, or meningoencephalitis were included. For disease progression probabilities, data not designating initial condition (e.g., death due to “listeriosis,” without designating CNS infection/sepsis) were excluded. Because disease progression data specific to cancer patients were lacking (studies included did not differentiate disease progression among different immunocompromised groups), these probabilities were based upon invasive listeriosis infections regardless of population subgroup. This decision is justified given the absence of cancer‐specific data and because the majority of listeriosis cases are in immunocompromised individuals (Almeida et al., 2010), a group that encompasses cancer patients.

The listeriosis risk distributions for eating a simulated salad consisting of RTE leafy greens, cucumbers, and tomatoes prepared by three strategies, (i) refrigeration only, (ii) rinsing and refrigeration as recommended by the FDA (FDA, 2020) (SFH diet), and (iii) surface blanching (applicable to cucumber and tomato only), and refrigeration, for which values were derived from a previous study (Gomez et al., 2023). Because the ND excludes raw salad ingredients, only noncompliant patients can develop listeriosis from salads. In this case, because compliance with the SFH diet is nearly 100% (Moody et al., 2018), it was assumed that those noncompliant with the produce restriction aspects of the ND were compliant with the FDA food safety guidelines (SFH diet). Therefore, P ₁ (the probability of acquiring listeriosis while on the ND) was equal to the product of ND noncompliance rate, hazard ratios (rates of general infection in cancer patients on the ND vs. SFH diet), and listeriosis risk on the SFH diet (Equation 5):

$$P_1=\mathrm{ND}\mathrm{noncompliance}\times \mathrm{SFH}\mathrm{risk}\times \mathrm{ND}\mathrm{hazard}\mathrm{ratio}.$$ (5)

A distribution for infection hazard ratios for the ND compared to SFH diet was created using data for general infection, fever, bacteremia, gastroenteritis, and blood stream infections in cancer patients assigned to either diet (Gardner et al., 2008; Jakob et al., 2021; Lassiter & Schneider, 2015; Moody, Finlay, et al., 2006; Tramsen et al., 2016; Trifilio et al., 2012). Data from three studies (DeMille et al., 2006; Lehrnbecher et al., 2008; Moody et al., 2018) were pooled to create a uniform distribution for patient compliance with the ND. Only one publication (Moody et al., 2018) studied compliance to the SFH diet, which was reported as 99.26% ± 4.8%. For simplicity, it was assumed that this compliance with the SFH diet was 100%. As compliance with a surface blanching diet has never been studied, it was assumed to be a uniform distribution from the median of the ND compliance distribution to 1, then adjusted in sensitivity analysis to determine the switchover point.

Parameters necessary for DALY and EV calculations of each outcome (Equations (1), (2), (3), (4)) were then characterized, beginning with outcome probabilities. For each probability of progression from gastroenteritis to hospitalization, and all other progressions to new illness states (e.g., sepsis to death and CNS infection to long‐term CNS complications), data from multiple sources were pooled into single distributions (Table 1). Probabilities of developing NEC due to the ND and SFH were acquired from the pilot study of Gupta et al. (2022). Rates of NEC for surface blanching and refrigeration only have never been published. In terms of produce inclusion and expected nutrient level, these diets are comparable to the SFH diet. Therefore, they were assumed to have the same rate of NEC as the SFH diet, and sensitivity analyses were later conducted to evaluate this assumption. In the absence of adult data, it was assumed that the pediatric NEC rate applied to all ages. NEC mortality rate differed for children and adults, so two separate distributions were fit to data from studies, including patients aged 0–25 (corresponding to study age groups 0–4 and 5–14, study indicated pediatric patients) and those older than 19 (corresponding to study age groups 15–44, 45–59, and 60+, study indicated adult population) (Abu‐Sbeih et al., 2020; Benedetti et al., 2021; Fiteni et al., 2018; Gil et al., 2013; Gomez et al., 1998; Gorschlüter et al., 2001; Gorschlüter et al., 2002; Lebon et al., 2017; Mokart et al., 2015; Seddon et al., 2018). Studies including both age ranges (e.g., patients aged 5–60) were excluded from these distributions.

TABLE 1.

Model input distribution parameters.

Parameter	Distribution	Source(s)
Listeriosis risk—all scenarios	Previously reported output vectors	Gomez et al. (2023)
Hazard ratio: ND vs. SFH	Laplace: location = 1, scale = 0.147889, truncated at 0	Gardner et al. (2008), Jakob et al. (2021), Lassiter and Schneider (2015), Moody et al. (2018), Moody, Finlay, et al. (2006), Tramsen et al. (2016), Trifilio et al. (2012)
Gastroenteritis duration (days)	Uniform: min = 1, max = 2.8	Gillespie et al. (2009), Goulet and Marchetti (1996), Scobie et al. (2019)
Probability hospitalization	Uniform: min = 0.918779, max = 0.95	Gillespie et al. (2006), Silk et al. (2012), The Institute of Environmental Science and Research Ltd. (2012)
Hospitalization (days)	Difference between uniform: min = 0, max = 20 and pre‐hospital duration	Rivero et al. (2003)
Gastroenteritis disability weight	0.393	Kemmeren et al. (2006)
Probability CNS infection	Log logistic: shape = 3.360269, scale = 0.325513, truncated at 1	Almeida et al. (2010), Büla et al. (1995), Doorduyn et al. (2006), Friesema et al. (2011), Gerner‐Smidt et al. (2005), Gillespie et al. (2006, 2009), Goulet and Marchetti (1996), Goulet et al. (2008), Guevara et al. (2009), Huang et al. (2010, 2011), Kiss et al. (2006), Koch and Stark (2006), Mook et al. (2012), Okutani et al. (2004), Rocourt et al. (1997), Scobie et al. (2019), Silk et al. (2012), Skogberg et al. (1992), Voetsch et al. (2007)
CNS infection duration (days)	Triangle: min = 1, likeliest = 21, max = 182	Arslan et al. (2015), Kemmeren et al. (2006)
CNS infection disability weight	0.426	de Noordhout et al. (2014)
CNS infection mortality rate	Uniform: min = 0.031250, max = 0.490909)	Amaya‐Villar et al. (2010), Aouaj et al. (2002), Arslan et al. (2015), Brouwer et al. (2006), Büla et al. (1995), Gil et al. (2013), Gillespie et al. (2009), Goulet and Marchetti (1996), Guevara et al. (2009), Kiss et al. (2006), Mook et al. (2011)
Probability long‐term CNS complications	Uniform: min = 0.118182, max = 0.20	Amaya‐Villar et al. (2010), Aouaj et al. (2002), Brouwer et al. (2006), Büla et al. (1995), Goulet and Marchetti (1996)
Long‐term CNS complications disability weight	0.292	de Noordhout et al. (2014)
Long‐term CNS complications duration	Estimated RLE—CNS infection duration
Probability sepsis	Uniform: min = 0.040146, max = 0.864967	Almeida et al. (2010), Büla et al. (1995), Doorduyn et al. (2006), Friesema et al. (2011), Gerner‐Smidt et al. (2005), Gillespie et al. (2006, 2009), Goulet and Marchetti (1996), Goulet et al. (2008), Guevara et al. (2009), Huang et al. (2010), Kiss et al. (2006), Koch and Stark (2006), Mook et al. (2012), Rocourt et al. (1997), Scobie et al. (2019), Silk et al. (2012), Skogberg et al. (1992), Voetsch et al. (2007)
Sepsis duration (days)	Uniform: min = 1, max = 7	Kemmeren et al. (2006)
Sepsis disability weight	0.210	de Noordhout et al. (2014)
Sepsis mortality rate	Uniform: min = 0.040146, max = 0.739130	Büla et al. (1995), Gerner‐Smidt et al. (2005), Gillespie et al. (2009), Guevara et al. (2009), Kiss et al. (2006), Mook et al. (2012)
Probability NEC, ND	0.33	Gupta et al. (2022)
Probability NEC, SFH, surface blanch, refrigeration only	0.05	Gupta et al. (2022), and assumptions
NEC disability weight	Uniform: min = 0.164, max = 0.348	Assumption based on Global Burden of Disease Collaborative Network (2020)
NEC duration (days)	Uniform: min = 1.5, max = 21.6	Abu‐Sbeih et al. (2020), Cartoni et al. (2001), Gil et al. (2013), Gomez et al. (1998), Shafey et al. (2013), Sloas et al. (1993), Sundell et al. (2012)
NEC mortality rate (adult)	Exponential: rate = 3.604231, truncated at 1	Abu‐Sbeih et al. (2020), Benedetti et al. (2021), Fiteni et al. (2018), Gil et al. (2013), Gomez et al. (1998), Gorschlüter et al. (2001, 2002), Lebon et al. (2017), Mokart et al. (2015), Seddon et al. (2018), Delignette‐Muller et al., (2023)
NEC mortality rate (child)	Exponential: rate = 15.93568	El‐Matary et al. (2011), Fike et al. (2011), McCarville et al. (2005), Moran et al. (2009), Mullassery et al. (2009), Rizzatti et al. (2010), Sahoo et al. (2022), Shafey et al. (2013), Sloas et al. (1993), Sundell et al. (2012)
ND compliance	Uniform: min = 0.57143, max = 0.92208	DeMille et al. (2006), Lehrnbecher et al. (2008), Moody et al. (2018)
SFH compliance	1	Assumption based on Moody et al. (2018)
Surface blanch compliance	Uniform: min = 0.57143, max = 1	Assumption
RLE 0–4 (years)	Uniform: min = 73.3, max = 77.8	Aries et al. (2021)
RLE 5–14 (years)	Uniform: min = 63.4, max = 73.3	Aries et al. (2021)
RLE 15–44 (years)	Weibull: shape = 6.269388, scale = 53.144223	Botta et al. (2019)
RLE 45–59 (years)	Logistic: location = 22.43124, scale = 3.20321, truncated at 0 and 45	Botta et al. (2019)
RLE 60+ (years)	Weibull: shape = 2.382438, scale = 12.367824	Botta et al. (2019)

Abbreviations: CNS, central nervous system; ND, neutropenic diet; NEC, neutropenic enterocolitis; RLE, remaining life expectancy; SFH, safe food handling.

The disability weight used for both non‐hospitalized and hospitalized gastroenteritis listeriosis patients was approximated as the value used by Havelaar et al. (2000) for Campylobacter gastroenteritis patients visiting general practitioners (GPs). This type of gastroenteritis was the more serious of the two described and thus more analogous to that faced by cancer patients and was used for both visiting GPs and hospitalized patients in a later assessment (Kemmeren et al., 2006). Disability weights for sepsis, CNS infection, and long‐term CNS complications were previously reported by de Noordhout et al. (2014), who derived these values by combining various states in the 2010 GBD study (Salomon et al., 2012). There is no published disability weight for NEC, so the WHO GBD study's disability weight range for “severe diarrheal disease,” acted as the bounds for a uniform distribution (Global Burden of Disease Collaborative Network, 2020). Symptoms corresponding with this condition are diarrhea ≥3 times a day, severe abdominal pain, nausea, and tiredness, which closely mimic symptoms of NEC (Abu‐Sbeih et al., 2020; Global Burden of Disease Collaborative Network, 2020).

Disease state durations were parameterized as described here forth. The distribution for duration of gastroenteritis prior to hospitalization was generated using the minimum and maximum mean times from symptom onset to hospitalization from three studies (Gillespie et al., 2009; Goulet & Marchetti, 1996; Scobie et al., 2019). The duration of hospitalization prior to progression to sepsis, CNS, or recovery was assumed to follow a uniform distribution for the largest range reported, which was for CNS infections, and encompassed the range for sepsis diagnosis (Rivero et al., 2003). Because the range was from symptom onset to specific diagnosis, the duration of illness prior to hospitalization was subtracted to obtain the duration of solely hospitalization. The durations of sepsis and long‐term CNS complications were assigned according to Kemmeren et al. (2006). A triangular distribution for duration of CNS infection was built using the minimum and median reported by Arslan et al. (2015), with the maximum being the point estimate reported by Kemmeren et al. (2006). Finally, because data were lacking, means and medians were both used to create a distribution for NEC duration. The distribution fit, range, and central tendency were reviewed to validate this assumption.

Data for cancer patients’ predicted RLEs based on age at diagnosis, for patients 0 and 1 years from diagnosis (Botta et al., 2019), were fit to distributions to obtain model RLE estimates. Ages utilized in the present study's age groups were as follows: (i) 17, 22, 27, 32, and 37 for ages 15–44; (ii) 45, 52, and 57 for ages 45–59; and (iii) 62, 67, 72, and 80 for ages 60+. When necessary, distributions were truncated at 0 and the RLE that, when summed with the minimum age in an age range, would equal a total lifespan of 90 years. Given that there were no data for cancer patients under 17, and maximum RLE for 17‐year‐old cancer patients approached that of the general population, it was assumed that for age groups 0–4 and 5–14, the RLE was a uniform distribution between the minimum and maximum CDC estimates for the ages within the group (Aries et al., 2021). Because data were reported in increments of five, RLE for 0 and 5, and 5 and 15, were used for the 0–4 and 5–14 age groups, respectively. For all health states, the duration of the previous health state(s) was subtracted from the RLE.

The results of the literature search and fit of input distributions to represent variability and uncertainty in each model parameter used to calculate model EVs are tabulated in Table 1.

2.6. Calculations

Input parameter distributions were fit to collected data in R version 4.0.5 (R Core Team), using the function “fitdist” from the package “fitdistrplus” (Delignette‐Muller et al., 2015). For parameters with ≥5 estimates, distributions were fit to data and ranked by the BIC statistic, with the best ranked distribution chosen. Occasionally, probability distributions included unfeasible values (<0 or >1). In such cases, distributions were truncated at 0, 1, or both, following fitting, using an ifelse function. Parameters with <5 estimates were assumed to follow a uniform distribution between the minimum and maximum estimates. Further calculations were executed in R for greater accessibility to statistical testing and manipulation.

EV equations were coded in R version 4.0.5 (R Core Team) (R Core Team, 2023). Monte Carlo analysis with 10,000 iterations and a set seed of 123 was used via Equation (4) to calculate each intervention's EV. Each iteration combined single values selected from probability distributions to produce 10,000 continuous DALY estimates. This resulted in median EVs being then compared to select the intervention strategies with the least DALYs. As a secondary analysis, sensitivity analyses of each decision tree were then conducted to determine switchover points—the value that justifies one strategy over the other strategy—for the most uncertain parameters.

2.7. Sensitivity analysis

One‐way analyses were conducted to understand the uncertainty and variability associated with the numerous input parameters within each decision model and how it affects the selection of the food safety diet. First, Spearman rank correlation coefficients were calculated for all applicable input parameter distributions and corresponding EV (cor.test function in R, with method specified as “spearman”).

Spearman rank correlation coefficients could not be calculated for point‐estimated or single‐value parameters. However, the authors surmised that some of these parameters were particularly uncertain, due to lacking literature (listeriosis risks resulting from the ND, SFH, surface blanching, and refrigeration only strategies, the probability of NEC associated with each intervention, compliance rates for SFH and surface blanching, and NEC disability weight) (Global Burden of Disease Collaborative Network, 2020; Gomez et al., 2023; Gupta et al., 2022), or variable due to inherent person‐to‐person variability (sepsis rate, CNS infection rate, and NEC durations and mortality rates), and should be included in a secondary sensitivity analysis. These parameters, and those identified as highly uncertain or variable via Spearman correlations, were further examined to determine their switchover point and the likelihood of their occurrence.

While calculating each switchover point, all values besides the parameter in question were held constant at a point estimate (the median of a parameter's respective distribution). Expected DALYs were then calculated and plotted for the ND and intervention in question as a function of the selected parameter, with the point of intersection signifying the switchover point.

3. RESULTS

3.1. DALY calculations and distributions

Across all interventions and age ranges, EVs, in DALYs, for SFH, surface blanch, and refrigeration interventions were nearly indistinguishable, varying only on the order of 10⁻⁴–10⁻³. As such, given SFH's high compliance rates (Moody et al., 2018) and resemblance to the current FDA and ACS recommendations (ACS, 2019; FDA, 2020), it is the sole alternative intervention discussed in detail here. Additionally, note that because this model framed the analysis in terms of the negative health outcomes (DALYs), larger EVs indicate worse health outcomes. EVs spanned from 1.8E−04 DALYs per person per chemotherapy cycle (SFH, 60+) to 24.5 (ND, 15–44) (Table 2). The ND resulted in the highest EVs, with median EVs at each age range up to seven times greater than for SFH. These results demonstrate that the expected negative health outcomes for ND are higher than they are for SFH for all age groups. For each intervention, the lowest and highest EVs occurred in the 60+ and 15–44 age groups, respectively. EV distributions were skewed left for both interventions and all age groups, with the ND consistently exhibiting a wider spread, which was attributable to the higher NEC risk on the ND (Figure 4). EVs in the 15–44 age group also displayed a wider spread than those of other age groups.

TABLE 2.

Expected value (disability‐adjusted life years [DALYs]) for each intervention and age group.

Intervention	Age	5%	50%	95%
Neutropenic diet	0–4	0.0875	1.0949	4.7000
	5–14	0.0792	0.9920	4.2798
	15–44	0.2222	3.1017	13.6571
	45–59	0.0977	1.3910	6.2722
	60+	0.0409	0.6185	3.1616
SFH	0–4	0.0125	0.1564	0.6712
	5–14	0.0114	0.1417	0.6112
	15–44	0.0317	0.4430	1.9504
	45–59	0.0140	0.1987	0.8958
	60+	0.0058	0.0883	0.4515
Surface blanch	0–4	0.0125	0.1564	0.6712
	5–14	0.0114	0.1417	0.6112
	15–44	0.0317	0.4430	1.9504
	45–59	0.0140	0.1987	0.8958
	60+	0.0058	0.0883	0.4515
Refrigeration only	0–4	0.0127	0.1565	0.6727
	5–14	0.0115	0.1421	0.6121
	15–44	0.0319	0.4431	1.9509
	45–59	0.0140	0.1989	0.8967
	60+	0.0059	0.0883	0.4515

Note: Higher DALYs indicate worse health outcomes.

Abbreviation: SFH, safe food handling.

FIGURE 4.

Expected value (EV) distributions for age groups 0–4 and 15–44.

Most (99.58%–99.95%) of the intervention EV magnitudes were attributable to EVs of YLLs associated with NEC. EVs of NEC YLDs accounted for a marginal fraction of the total (0.04%–0.23%), and EVs of listeriosis YLDs and YLLs were a negligible contributor. A sample categorization of median EVs by DALY component and outcome is shown in Table 3.

TABLE 3.

Median expected values categorized by disability‐adjusted life year (DALY) component and outcome, neutropenic diet (ND) 15–44.

YLD LM ill (%)	YLD LM hosp (%)	YLD sept (%)	YLD CNS inf (%)	YLD CNS sequelae (%)	YLD NEC (%)	YLL sept (%)	YLL CNS inf (%)	YLL NEC (%)
6.22E−13 (∼0)	2.69E−12 (∼0)	3.08E−12 (∼0)	1.01E−11 (∼0)	2.27E−10 (∼0)	2.56E−03 (0.08)	1.71E−09 (∼0)	1.02E−09 (∼0)	3.10E00 (99.92)

Abbreviations: CNS, central nervous system; NEC, neutropenic enterocolitis; YLL, years of life lost.

3.2. Sensitivity analysis

For all age groups and both interventions, the parameter with the highest Spearman correlation coefficient was NEC mortality rate, decreasing with increased age and ranging from 0.90 to ∼1 (0.9998). The second‐most influential parameter and its relationship strength to an intervention's EV depended on age group. For the 0–4 age group, this was the listeriosis hazard ratio for ND versus SFH, with a low correlation coefficient (0.02). For all other age groups, RLE had the second‐highest correlation coefficient, with value increasing by age: 0.03 (5–14), 0.16 (15–44), 0.22 (45–59), and 0.39 (60+). The absolute values of the remaining parameters’ correlation coefficients were ≤0.03.

3.3. Switchover points

Negative health outcomes, as indicated by EVs in DALYs, were lower for the SFH diet than for the ND in all plausible scenarios (within reported ranges). Under all conditions, switchover points for ND and alternative intervention NEC probabilities occurred when both food safety diets had the same NEC probability of either 0.05 or 0.33. The ND was only preferred when SFH NEC probability was greater than 0.33−1×10⁻⁰⁹. This reveals that listeriosis outcomes do not influence the model decision unless the difference in intervention NEC probabilities is within 1E−09. Intervention EVs as a function of NEC probabilities, for ND and SFH, ages 15–44, are shown in Figure 5. Varying slightly by age group, the SFH diet listeriosis risk switchover points ranged from 2.9E−02 (SFH, ages 45–59) to 2.8E−01 (SFH, ages 15–44) (Table 4), indicating that the ND is only justified when listeriosis risk associated with the SFH diet is remarkably high (compared to previous estimates, which range from 8.4E−13 to 2.5E−06) (Ding et al., 2013, Gomez et al., 2023, and FDA, 2003). Alternatively, all switchover points for ND listeriosis risk were outside the range of possible values for risk (0–1), ranging from −2.96E−01 to −2.95E−02, characterized by age groups. This signifies that feasible listeriosis risk reductions attributable to the ND do not justify its use over other interventions.

FIGURE 5.

Switchover points for neutropenic enterocolitis (NEC) probability for neutropenic diet (ND) and safe food handling diet (SFH), ages 15–44.

TABLE 4.

Listeriosis risk switchover points for alternative interventions.

Intervention	Age	Switchover point
Safe food handling	0–4	0.0654
(SFH)	5–14	0.0633
	15–44	0.2841
	45–59	0.0290
	60+	0.2836
Surface blanch	0–4	0.0568
	5–14	0.0550
	15–44	0.2626
	45–59	0.0250
	60+	0.2620
No intervention	0–4	0.0496
	5–14	0.0480
	15–44	0.2285
	45–59	0.0218
	60+	0.2280

Abbreviation: SFH, safe food handling.

Figure 6 shows EV as a function of NEC adult mortality rate and RLE, respectively. Despite being identified as influential parameters in the sensitivity analysis, and EV increasing dramatically with each parameter, these parameters had implausible (negative) switchover points. All other parameters also had implausible switchover points (e.g., a probability greater than one and an illness duration of many or negative years). As such, realistic changes in aforementioned parameters did not influence the preferred dietary intervention.

FIGURE 6.

Neutropenic enterocolitis (NEC) adult mortality rate and relative life expectancy (RLE) switchover points for the neutropenic diet (ND) versus the safe food handling diet (SFH), ages 15–44.

4. DISCUSSION

This decision model is the first to estimate and compare the negative health impacts of listeriosis and the negative health impacts from oft recommended diets for cancer patients as related to fresh produce. Previous works calculated DALYs for listeriosis from all foods in the general population (de Noordhout et al., 2014; Havelaar et al., 2012; Scallan et al., 2015), which cannot be directly compared with DALYs for the target population. For instance, authors used standard life expectancies for all age groups, which are longer than those expected for adult cancer patients, driving up YLLs. Additionally, cancer patients’ increased listeriosis risk is likely to increase expected DALYs beyond that of the general population. Further, no existing studies impute cancer patients’ NEC disease burden in DALYs. Present study NEC EVs were high (median DALYs per person per chemotherapy cycle ranged 0.095–3.018 per person); however, the present study is specific to cancer patients, who are much more susceptible to severe intestinal disease, and have a high individual NEC mortality rate (median 18% in adults). Additionally, the literature did not provide NEC rates in adult cancer patients, only children. Although switchover point analysis revealed that a different NEC rate is unlikely to change the preferred decision, this information would improve the accuracy of the model. Therefore, although this model has provided the first estimates of listeriosis and NEC EVs, more data on burden of disease specific to cancer patients are necessary for validation. Intervention‐specific EVs were highest for the ND and virtually indistinguishable among the other interventions. Although any of these interventions may be recommended over the ND, it is prudent to consider ease of administration and diet compliance. Compliance with a surface blanching diet has never been tested, yet it is reasonable to anticipate that administration may be difficult, and adherence viewed as tedious and time‐consuming, resulting in subpar compliance rates. Although the refrigeration only diet may technically be the easiest to follow, it contradicts most hospital recommendations, and over 80% of cancer patients report thoroughly washing fruits and vegetables before eating most to all of the time, which hints to an underlying habit or desire to practice this food safety behavior (Paden et al., 2019). In contrast, the SFH diet is presently recommended by the FDA (FDA, 2020), and current ACS guidelines (2019) closely match. SFH guidelines are familiar, simple to follow, and have a published compliance rate nearing 100% (Moody et al., 2018). For these reasons, the SFH diet is currently recommended, although acceptability and barriers to compliance among cancer patients will require further examination.

The decision model's EVs were overwhelmingly dominated by NEC YLLs. This can be attributed to the differences in risk between NEC (0.33 for ND and 0.05 for the SFH diet) and listeriosis (medians ranging from 7.21E−10 to 4.38E−08). When combined with a high NEC mortality rate, this imbalance led to NEC outcomes driving decision making and listeriosis outcomes having an inconsequential effect on model results. Because the SFH diet had a lower NEC risk than the ND, it was always preferred to the ND. In a secondary analysis, the model was run with 100,000 simulations to determine if these results stabilized. On average, EVs differed only on the order of 10⁻⁴ to 10⁻³, with preferred decisions and switchover points remaining the same, indicating stabilization with 10,000 iterations.

Adjusting listeriosis parameters, like probability and duration of CNS or sepsis, within a conceivable range did not change the preferred intervention from the SFH diet to the ND. Furthermore, switchover point analysis revealed that for listeriosis outcomes to become quantitatively meaningful, median listeriosis risk on the SFH diet would have to increase to a range of 2.9E−02 to 2.8E−01. When previously reported vegetable‐specific listeriosis risks (Ding et al., 2013; FDA, 2003; Franz et al., 2010; Sant'Ana et al., 2014) were compounded for the duration of a chemotherapy cycle (ACS, 2019), the 95th percentiles ranged from 4.00E−08 to 9.07E−05, which is three to four orders of magnitude lower than the switchover point risk. Additionally, the switchover points for ND RTE salad listeriosis risk were all negative, indicating that even if a patient perfectly complied with the produce exclusion aspects of the ND, thereby reducing listeriosis risk from RTE salads to zero, it would not be enough to counteract the ND's increased NEC risk, and EV would still be highest on the model's ND arm. These findings emphasize the improbability of reaching a listeriosis risk that would justify the use of the produce restrictive aspects of the ND over the SFH diet and imply that the risk–risk tradeoff of negative secondary outcomes, such as NEC, should be the main concern for any food safety diet.

The sensitivity analysis determined that NEC mortality rate had a nearly perfect positive relationship with intervention EV (Spearman rank correlation coefficients 0.90 to ∼1). This is another manifestation of NEC's dominance in the model, considering NEC mortality rate plays a large role in determining EVs of NEC YLLs. The high correlation coefficient may also reflect the fundamental variability in disease mortality rates, which depend on individual health, age, and treatment options. An attempt was made to mitigate age‐related variability using two distinct mortality rates (child and adult). However, many of the adult mortality studies included a large age range (e.g., ages 21–78), and were still intrinsically variable. In the future, public health experts may find it useful to divide this age range to collect more accurate NEC mortality rate data. Nonetheless, despite this variability, NEC mortality rate switchover points did not occur in a feasible range, suggesting that better characterized variability would not change model recommendations.

As the second major component of NEC YLLs, RLE had the second‐highest correlation coefficient for all age groups, excluding ages 0–4. Again, this exemplifies the strong influence NEC outcomes had on model EV. Similar to NEC mortality rate, the high correlation coefficient underscored the inherent variability of cancer patients’ life expectancies, which can differ drastically by individual health and lifestyle, cancer type, age at diagnosis, and treatment options. Strategies used to characterize this variability included constructing the RLE distribution based on data for numerous cancer types and diagnosis age, as well as separating models by age group, though fluctuations will also remain. Further, the switchover points for NEC mortality rate were all negative, once again establishing that plausible variations do affect model recommendations.

4.1. Model limitations

This study's key finding was that the risk and outcomes of NEC markedly outweighed those of listeriosis, and because the ND had a higher NEC risk than other food safety diets, its use is not justified based on the analyses conducted in this study. It is important to consider that the only available data on diet‐based NEC risk were sourced from a pilot study, with 21 children each on ND and “standard diet” (equivalent to SFH diet) arms (Gupta et al., 2022). Even if a full‐scale study found a smaller disparity among diet‐based NEC risks, the model results would remain the same because SFH diet NEC probabilities must be greater than or equal to those of the ND to justify the ND. Therefore, a smaller discrepancy would not change the preferred decision. Nonetheless, collecting more data to better represent diet‐based NEC risk is crucial for enhancing model accuracy.

Despite model recommendations not fluctuating by age group, the groupings could provide the opportunity to define inherent variability, if selected carefully. Being that model EVs were dictated by NEC YLL, NEC mortality rate and RLE are key indicators for choosing effective age groupings. Model EVs were similar for the two youngest age groups, peaked in the 15–44 age group, and decreased dramatically in the 60+ age group. Age groups 0–4 and 5–14 were analogous because they utilized the same NEC mortality rate and had similar RLE distributions. The 15–44 age group EV spike can be ascribed to the combination of the high adult NEC mortality rate and wide‐ranging RLE distribution (range 42 years), which produced artificially high EVs (e.g., a 44‐year‐old with adult NEC mortality and RLE of 70). This is also why this age grouping's EV distribution was markedly widespread (Figure 4). EVs for those aged 45–59 were decreased in correspondence with the lower, more accurate, RLE. Because RLEs for the 60+ age group were considerably lower than for others, this grouping had the lowest EVs. Based on these inferences, it may be more accurate to have three age groups based on NEC mortality rate and RLE: (i) children and young adults (≤30), (ii) adults (31–59), and (iii) older adults (60+).

It was not possible to quantify some effects of the ND for use in this model. Previous works documented negative mental health and quality of life effects in cancer patients assigned to restrictive diets (McGrath, 2002; Moody, Meyer, et al., 2006). Unfortunately, these effects have never been clearly defined, and there are no data regarding their prevalence on ND and alternative intervention diets. Additionally, given that the ND provides inadequate fiber and vitamin C compared to the SFH diet (Maia et al., 2018), it is possible that it may lead to deficiencies in compliant patients (Moody et al., 2002). Nutritional deficiencies during cancer treatment can worsen cancer prognoses, quality of life, and cause treatment complications (Haskins et al., 2020). Including DALYs associated with these effects could drive the ND EVs even higher, making preference for the ND even more unlikely.

For simplicity, this model considered two health outcomes, (i) listeriosis and (ii) no listeriosis, with the probability of NEC. However, there is a very small probability (6.85E−11 to 1.05E−10) that both disease states could occur simultaneously. In this case, the combined disability weight would likely increase in correspondence with the presumed more severe form of illness, and duration would likely lengthen. Despite this, the combined outcome would not be expected to influence model results, given the event's low likelihood.

Finally, this model is applicable to listeriosis and the consumption of RTE salad products only, but the ND eliminates the consumption of all fresh produce to minimize risk from multiple pathogens, including nontyphoidal Salmonella and Shiga toxin–producing Escherichia coli O157:H7. Listeriosis was chosen as the model's focus, given that it primarily affects the immunocompromised (e.g., cancer patients) and has the highest mortality rate of the three pathogens (Scallan et al., 2011). Because Salmonella and E. coli O157:H7 also carry substantial burdens of disease (Scallan et al., 2015), incorporating risk assessments for these pathogens, in addition to L. monocytogenes, in a variety of produce would broaden the decision tool. In the future, it may also be useful to collect consumption and disease progression data necessary to account for other immunocompromised groups who face an increased risk of listeriosis, such as transplant patients.

Finally, human decision making, especially during cancer treatment, is affected by a myriad of variables, given the vast range of experiences with treatment and life circumstances. Further, current risk‐informed educational materials are incongruous and vary from institution to institution (Braun et al., 2014; Carr & Halliday, 2015; French et al., 2001; Smith & Besser, 2000). For the information presented in this study to translate to patients, strategic, streamlined communication interventions must be considered.

5. CONCLUSION

The model presented in this study is the first to quantify cancer patients’ health outcomes associated with the ND and less disruptive alternatives. Due to the predominant effects of NEC, which was more prevalent on the ND, the three alternative food safety diets were preferred over the ND produce guidelines in all age groups. Of the three, the SFH diet is recommended due to its high compliance rates and ease of implementation, though future work should evaluate acceptability and compliance barriers for cancer patients. Estimated listeriosis EVs were comparable to previous reports, but the lack of NEC disease burden data made full model validation difficult. More data are needed to ensure accuracy in this area. Switchover point analysis revealed that for all parameters but diet‐based NEC risk, there were no plausible scenarios favoring implementation of ND produce restrictions. Further investigation is needed to ensure this parameter's accuracy, as current data are from a pilot study. Most importantly, this study provides quantitative evidence supporting the adoption of produce‐inclusive food safety diets and discontinuation of ND produce restriction recommendations during cancer treatment.

Gomez, C. B. , Mitchell, J. , & Marks, B. P. (2025). A decision analysis of cancer patients and the consumption of ready‐to‐eat salad. Risk Analysis, 45, 1282–1297. 10.1111/risa.17658