Inverse association between Paleolithic Diet Fraction and mortality and incidence of cardiometabolic disease in the prospective Malmö Diet and Cancer Study

The MDCS is a population-based prospective cohort study conducted in Malmö, a city in the south of Sweden. Baseline examinations were conducted between 1991 and 1996. All women born during the period 1923–1950 and all men born 1923–1945, which were living in the city of Malmö, were invited to participate (n = 74,138) in the study. Details of the cohort and the recruitment procedures are described elsewhere [9]. The only exclusion criteria were absence of mental capacity and inadequate Swedish language skills (eligible persons = 68,905). The participants filled out questionnaires covering socioeconomic, lifestyle, dietary factors, and recorded meals, and underwent a diet history interview. Anthropometric measurements were conducted by nurses. Weight was measured using a balance-beam scale with subjects wearing light clothing and no shoes. Standing height was measured with a fixed stadiometer calibrated in centimeters. Waist circumference was measured midway between the lowest rib margin and the iliac crest. During the screening period, a total of 28,098 participants (41% of the eligible persons) completed all baseline examinations (Supplemental Fig. 1). Of these, 2129 participants with baseline examinations in 1991 were excluded from the present study since the coding of vegetable and fruit subgroups were handled differently for the 1991 database — compared with later — thus precluding calculation of PDF with legumes excluded from other vegetables as non-Paleolithic [10]. A further 1865 participants with diabetes (n = 1177), a history of coronary events (n = 507) and/or stroke (n = 290) at baseline were also excluded. The remaining 24,104 participants who completed all baseline examinations were included in the present study. All study participants gave written informed consent, and the study was approved by the Regional Ethics Review Board in Lund, Sweden (Dnr LU51-90).

Dietary data were collected once at baseline. The MDCS used an interview-based, modified diet history method that combined (I) a 7-day food record of meals that varied from day to day, usually lunch and dinner meals, cold beverages, and nutrient supplements, and (II) a 168-item food frequency questionnaire for the assessment of consumption frequencies and portion sizes of regularly eaten foods that were not covered by the 7-day food record. Finally, (III) during a 45-min interview, questions were asked on usual portion sizes and cooking methods of foods recorded during the 7-day food record. The MDCS method is described in detail elsewhere [11, 12]. The PDF association analyses were adjusted for a binary variable called “diet method version” because slightly altered coding routines of dietary data were introduced in September 1994 to shorten the interview time (from 60 to 45 min). The altered coding routines resulted in two slightly different method versions (before or after September 1994) without any major influence on the ranking of individuals [12]. The validity of the MDCS method was evaluated in the Malmö Food study 1984–1985, comparing the method with 18 days of weighed food records [13, 14]. The Pearson correlation coefficients, adjusted for total energy intake, between the reference method and the MDCS method, were in women and men, respectively, for intakes of bread 0.58/0.50, cereals 0.73/0.74, fruits 0.77/0.60, vegetables 0.53/0.65, low-fat milk 0.92/0.90, high-fat milk 0.75/0.76, cheese 0.59/0.47, fish 0.70/0.35, low-fat meat 0.51/0.43 and high-fat meat 0.80/0.40 [13].

The mean daily intake of foods was calculated based on frequency and portion size estimates from the questionnaire and food record. The food intake was converted to energy and nutrient intakes using the MDCS nutrient database where the majority of the nutrient information comes from PC-KOST2-93 from the National Food Agency in Uppsala, Sweden. Energy intake was divided into four categories based on age and sex-specific reference values from the Nordic Nutritional Recommendations (NNR2012) for energy intake at low, medium, and high physical activity level (< low, low-medium, medium–high, > high) [15]. Evaluations of energy and nutrient intakes were based on the Nordic Nutritional Recommendations (NNR2012) [15]. The food intakes in MDCS were aggregated into 33 groups to obtain food groups more frequently consumed in the population, but to keep characteristics related to both dietary behaviors and nutrient content.

Briefly, as previously described, Paleolithic diet is specified on a basis of included and excluded food groups. It includes fruits and vegetables, roots and tubers, lean meats, eggs and nuts but excludes grains, dairy products, legumes, refined fats and sugar [5]. The PDF is calculated as the mean daily fraction between the summed intake of all food belonging to the Paleolithic food groups and the summed intake of all food [5]. For the present study, non-energy-containing beverages such as water, coffee and tea were excluded, and the 33 MDCS food groups were further aggregated into food groups comparable to previous RCTs with PDF (Supplemental Table S1). Among these food groups, which were deemed consistent with our previous classification as Paleolithic in calculations of PDF, were ‘vegetables’, ‘fruits’, ‘potatoes’, ‘eggs’, ‘meat’ (pork, beef, lamb, game meat, poultry, and pure offal), ‘fish’, ‘oil rape seed and olive’, ‘nuts’, and ‘wine’ [5, 6] (Supplemental Table S1). The remaining further aggregated food groups were consequently classified as non-Paleolithic, and consisted of ‘legumes’, ‘juice’, ‘meat products’ (offal as a mixed product or spread, and sausage), ‘milk and milk products’, ‘sweet beverages’, ‘cereals with rice’, ‘fat oil and margarine’, ‘bakery sweets’, ‘jam’, ‘sauce soups’, ‘beer’, ‘spirits’, and ‘remainder miscellaneous’ [5, 6] (Supplemental Table S1). PDF was calculated by weight as the fraction of the mean daily summed absolute dietary intake of Paleolithic food groups divided by the mean daily summed absolute dietary intake of all food groups, as described previously [5, 6] (Supplemental Table S1).

The participants contributed person-time from the date of enrolment until date of diagnosis, death (10,092 participants), migration from Sweden (188 participants), or end of follow-up (December 2019), whichever occurred first. The mean follow-up time was 20 years for type 2 and unknown type diabetes and 21 years for mortality, coronary events, and ischemic stroke event (range 0–28 years for all).

The National Tax Board provided information on vital status and emigration. Cause of death was obtained from the Cause of Death Registry, where International Classification of Diseases (ICD) codes for the underlying main cause of death were registered. The ICD codes from version 9 and 10 used to record main cause of death group for participants were as follows; tumor (ICD9:140–239, ICD10:C,D00-D48), neurological disease (ICD9:320–389, ICD10:G,H), cardiovascular disease (ICD9:390–459, ICD10:I), respiratory disease (ICD9:460–519, ICD10:J), and digestive disease (ICD9:520–579, ICD10:K). Validation of mortality from the cause of death register ICD codes found that cardiovascular mortality was confirmed in 94% of the participants [16].

Participants with at least two HbA1c values above 6.0% with the Swedish Mono-S standardization system (corresponding to 6.9% in the US National Glycohemoglobin Standardization Program and 52 mmol/mol with the International Federation of Clinical Chemistry and Laboratory Medicine (IFCC) units) [17, 18] were categorized as diabetes cases in the Malmö HbA1c Registry. In addition, diabetes cases were identified via four registries from the National Board of Health and Welfare in Sweden: The Swedish National Inpatient Registry, the Swedish Hospital-based Outpatient Care Registry, the Cause of Death Registry, and the Swedish Prescribed Drug Registry. When available from the regional Diabetes 2000 registry of Scania and its successor the All New Diabetics In Scania (ANDIS) registry, type of diabetes was based on the glycemic parameters, treatment/medication, age at diagnosis, glutamic acid decarboxylase antibodies (GADA), C-peptide and body mass index (BMI). Included were cases with type 2 diabetes and unknown type diabetes. The majority of cases with unknown type diabetes cases were assumed to have type 2 diabetes. Excluded were cases with type 1 diabetes, Latent Autoimmune Diabetes in Adults (LADA), pregnancy diabetes, secondary diabetes, and other types of diabetes. If available, we used information on the date of diabetes diagnosis from two registries prioritized in the following order: (a) the Diabetes 2000 registry of Scania and (b) the Swedish National Diabetes Registry. These registries required a physician diagnosis according to established diagnosis criteria (fasting plasma glucose concentration ≥ 7.0 mmol/L or fasting whole blood concentration ≥ 6.1 mmol/L, measured on two different occasions).

Information about prevalent and incident coronary and ischemic stroke events was taken from the national Swedish Hospital Discharge register, Cause-of-death register, and the local stroke register in Malmö (STROMA). A coronary event was defined based on codes 410–414 (fatal or non-fatal myocardial infarction or death due to ischemic heart disease) in the International Classification of Diseases, 9th Revision (ICD-9). Ischemic stroke was defined based on code 434 (ICD-9) and diagnosed when computed tomography, magnetic resonance imaging or autopsy could verify the infarction and/or exclude hemorrhage and non-vascular disease. If neither imaging nor autopsy was performed, the stroke was classified as unspecified. Hemorrhagic and unspecified stroke cases (ICD-9 code 430, 431 and 436) were excluded since these subtypes of stroke do not have the underlying risk factors associated with diet as ischemic stroke has.

Information on age was obtained from the personal identification number, which is assigned to individuals upon birth or permanent immigration to Sweden. BMI was calculated from measurements of weight and height and divided into four categories(< 18.50, 18.50–24.99, 25–29.99, ≥ 30.00 kg/m²) [19]. Leisure time physical activity was assessed by asking the participants to estimate the number of minutes per week spent on 17 different activities. The duration was multiplied with an activity-specific intensity coefficient and an overall leisure time physical activity score was created and divided into quintiles [16]. The smoking status of the participants was defined as current smokers (including irregular smokers), ex-smokers, and never-smokers. The calculated mean daily alcohol intake from the 7-day food record was divided into quintiles by sex. Participants were divided into four categories according to their highest level of education (≤ 8, 9–10, 11–13 years, or university degree). Season was defined as the season of diet data collection (winter, spring, summer, and fall/autumn).

The SPSS statistical computer package (version 28.0; IBM Corporation, Armonk, NY, USA) was used for all statistical analyses. Normality was assessed both graphically and numerically including Kolmogorov–Smirnov tests and many variables were non-normally distributed with outliers. Baseline characteristics were examined using Mann–Whitney U test, Chi-square test and Spearman rank test. Cox proportional hazards regression model was used to examine associations between the predictor PDF (both continuous and by quintiles) and risk of the outcomes all-cause death, main cause of death group, coronary event, ischemic stroke, incidence of type 2 diabetes, and incidence of either type 2 or unknown type diabetes. PDF as a continuous variable varies between 0 and 1. With PDF as a continuous variable, the hazard ratio indicates the change in the risk if PDF rises by one unit from 0 to 1, thus comparing a PDF of 0% with a PDF of 100%. For every 10% difference in PDF the percentage difference in risk will then be 10 multiplied with the difference between 1 and the hazard ratio. Years of follow-up was used as the underlying time variable. Predictor selection for our Cox regression model was based on our inferential goal of evaluating a predictor of primary interest, namely PDF. In pursuing this goal, relatively inclusive models are more likely to minimize the central problem of confounding. Predictors necessary for face validity as well as those that behave like confounders should be included in the model whilst avoiding predictors that are alternate measures of or mediators of our predictor of primary interest [20]. The full Cox regression model included the following predictors obtained from baseline examination: PDF (continuous or categorical), age (continuous), sex, diet method version, season, leisure time physical activity, smoking, alcohol intake, education, BMI, energy intake, born in Sweden, and living alone (categorical). The first PDF quintile was used as the reference, indicating participants with the lowest PDF. The continuous variables BMI and energy intake were converted into categorical predictors for the Cox regression model due to violation of its linearity assumption. The predictors were identified from the literature and indicated potential confounding in the MDCS cohort due to associations with examined outcomes and dietary intakes. Effect measure modification by predictors was assessed with stratification and interaction terms in Cox regression analyses. Since associations between PDF and examined outcomes could be mediated via the effects of PDF on BMI and energy intake, we also performed Cox regression analyses with BMI and energy intake excluded from the full model. Missing values for the predictors of the full model Cox regression were treated as separate categories. In sensitivity analyses, we excluded participants with missing values for predictors of the full model Cox regression. To assess the proportional hazards assumption, log minus log graphs were used to test interactions between the underlying time variable and examined categorical predictors (age categorized into three age categories 44–54, 55–64 and 65–74 years for this assessment). All statistical tests were two-sided and statistical significance was assumed at p < 0.05.

A strength of this study is its large sample size and long follow-up period of 20 years on average. The MDCS dietary assessment method measures both habitual and recent intakes, and the documented relative validity and reproducibility of food intake indicates high-quality dietary data. Moreover, it is a population-based prospective study, which reduces the risk of selection bias and reverse causation. Further strengths are the extensive information on potential confounders, and that diet was measured with a modified diet history method including a 7-day food record for cooked meals.

A major limitation of this observational study is that diet and other lifestyle factors were only measured at baseline, precluding the possibility of studying and adjusting for changes over time. However, reproducibility studies including similarly aged participants show acceptable agreement between repeated dietary measurements [53, 54], possibly because dietary habits are often already established at younger ages. For other lifestyle factors, the impact of this limitation can be expected to vary. For BMI, both decreases and increases over time are associated with greater risk of cardiovascular outcomes [55] indicating that the risk that is based on a single measurement at baseline could be underestimated. Changes in leisure-time physical activity are inversely associated with changes in cardiovascular outcomes [56]. Leisure-time physical activity in Sweden have increased during the time period of this study which indicates that the associated risk from a single measurement at baseline could be overestimated [57, 58]. Decreased smoking prevalence and increased alcohol intake among the general population in Sweden during the time period of this study indicate that the associated risk from a single measurement at baseline could be overestimated and underestimated, respectively [59, 60].

Another limitation is that the study sample only consisted of Swedish individuals aged 44–74 years and mostly women (63%) from an urban setting, which decreases generalizability to other age groups, men, and rural settings. Furthermore, residual confounding cannot completely be ruled out despite adjustment for several confounders, and changes over time in recognition, diagnosis, certification and/or coding of different diseases may have affected the registered death causes and consequently also the found associations. A reflection of this may be the inconsistent trends in hazard ratios across quintiles of PDF seen for several cause-specific mortality outcomes, requiring greater caution in their interpretation. In addition, since one of our main objectives was to examine all-cause mortality, we did not take competing risks into account in our statistical analysis, hence why we cannot exclude that this may be a concern regarding the results on cause-specific mortality. Finally, a limitation of this study is that other dietary estimates and pattern scores were not calculated, precluding the possibility of direct comparison with PDF. Such comparisons, albeit not within the planned scope of this study, would be interesting and are considered for future studies.

The results suggest the clinical importance of PDF which warrants further study. Future dietary research should assess PDF whenever possible to benefit from the comparability of PDF between studies.