A non-invasive modifiable Healthy Ageing Nutrition Index (HANI) predicts longevity in free-living older Taiwanese

Nutritional factors contributing to disability and mortality are modifiable in later life. Indices would add utility. We developed a gender-specific Healthy Ageing Nutrition Index (HANI) for all-cause mortality in free-living elderly. We stratified 1898 participants aged ≥65 y from the 1999–2000 Nutrition and Health Survey in Taiwan by region and randomly allocated them into development and validation sets. Linkage to the National Death Registry database until December 31, 2008 enabled mortality prediction using Cox proportional-hazards models. Four factors (appetite, eating with others, dietary diversity score, and BMI) with best total of 25 HANI points for men; and 3 factors (cooking frequency, dietary diversity score, and BMI) with best total of 27 HANI points for women, were developed. In the validation set, the highest HANI group exhibited a greater intake of plant-derived food and associated nutrients, a favourable quality of life, and more muscle mass, compared with the lowest group. The highest HANI group predicts mortality risk lower by 44 percent in men and 61 percent in women. Adjusted mortality HRs were comparable between sets. HANI is a simple, non-invasive, inexpensive, and potentially modifiable tool for nutrition monitoring and survival prediction for older adults, superior to its individual components.

For the entire cohort, the HRs for men with HANIs of 14-16 and >16 were 0.50 (0.36-0.70) and 0.35 (0.24-0.52), respectively, compared with those with HANI < 14. Women with the highest HANI exhibited a 70% lower mortality risk (HR: 0.30, 95% CI: 0.17-0.54). The entire study cohort behaved in the same manner as the development set, as shown in the survival curves in Fig. 3. There was equally good discrimination between the development and validation sets. Men in the validation set had a similar and overlapping cumulative survival rate between HANI 14-16 and that greater than 16. This is confirmed by comparable Harrell's C and Somers' D scores, more so for women than men.

Discussion
It is possible to identify factors that are non-invasively obtained and are potentially modifiable nutrition-related predictors of survival. The gender-specific and composite indices for survival prediction were the dietary pattern, food preparation, social circumstances of eating, and body composition. Confidence in HANI was gained by the use of predictive power statistics applied to both development and validation data sets. Moreover, the predictability of HANI is superior to that of any of its individual components.
The nutrition-related factors that are mainly associated with reduced survival are less PA 10 , loss of appetite 21 , chronic energy deficiency 30 , weight loss 31 , and sarcopenia 32,33 , and all are interconnected. In a Taiwanese cohort, regular moderate PA for 30-40 min daily was associated with reduced mortality 10 . However, PA did not modify the association between HANI and survival, suggesting that there may be another means of formulating a predictor of survivorship. In this study, we may have captured potentially predictive variables like PA and frailty by the inclusion of others like eating alone, cooking or BMI. The Mini Nutritional Assessment (MNA) is used to detect institutional undernutrition; in its short-form (6 items) it is not associated with mortality, although each item alone is associated with mortality in older free-living women 34 . In HANI for men, loss of appetite, as detected by the MNA, is an individual predictor of mortality. The Geriatric Nutritional Risk Index comprises weight, weight loss, and albumin which limits its use in community settings because of the dependence on memory and the need to obtain biomarkers 35  Measured by a 24-hour dietary recall of 6 food groups or by asking the following question. "Did you eat more than half a serving size of any of the following foods yesterday?" • Breads, cereals, starches (e.g., bread 1 slice, cereal 1/2 cup, bagels 1/2, white rice 1/2 cup) • Dairy (e.g., milk 1/2 cup, yogurt 50 g, cheese 1 slice) • Meat, fish, egg or legumes (e.g., 1.5 oz cooked meat, egg 1/2, soymilk 1/2 cup) • Vegetables (e.g., 1/2 cup) • Fruits (e.g., oranges 1/2, apples 1/2, pears 1/2, bananas 1/2) • Fats and oils (e.g., 1/2  36 . However, the predictive ability of this index for disability, disease, and mortality is unclear. In HANI, DDS and BMI are common to both genders by analysis and deduction. Encouragement to improve dietary quality exerts favourable effects on food patterns and the intake of nutritious food components. The shared gender relevance of a diverse diet in survival is evident 19,37 . DDS was measured by a 24-hr dietary recall in this study. It can be rapidly assessed by asking participants if they consumed half a serving size of each food group on the previous day (such information can be inserted into the online HANI app https://ychuang.shinyapps. io/HANI/, Supplementary Fig. S1). This can be applied in both community and clinical settings for nutritional evaluation and education.
The finding presumably reflects the importance of adequate energy throughput to the achievement of an adequate intake of favourable food components (possibly with more fat mass), along with preservation of muscle (less sarcopenia). As in other studies, we identified a higher BMI to be a survival advantage. The waist circumference was not an independent survival predictor. Sarcopenia increases with age and is often obscured by increasing body fat 33 . Along with reduced muscle strength, it contributes to frailty and mortality. Similar to sarcopenia, SMMI is a predictor of survival in this population 32 and positively associated with HANI. SMMI has been taken into account for this population, used in the cross-validation of HANI, but are not included in the indices as items which are less routinely available. Good appetite was a survival advantage in men, but not in women. The gender difference may be based on the relative inability of men to maintain healthy dietary and PA practices, which encourage a more appropriate appetite with age. The corollary in women, to their advantage, would be that more frequent cooking contributes substantially to the association of HANI with survival 18,38 . We hypothesize that the findings related to the predictive ability of HANI for survival (i.e., appetite and eating alone in men and cooking in women) are linked. Anorexia associated with ageing and loss of desire to eat are contributors to poor nutritional status 21 . Pathophysiologic anorexia associated with ageing develops when there is failure to regulate food intake adequately 39,40 . In the NHANES III of America, food intake decreases linearly by age and is probably associated with reduced PA and energy metabolism 40 . However, increased energy intake was associated with increasing HANIs in men (P = 0.006, data not shown), but not in women. This finding suggests that energy intake and higher HANI are more dependent on appetite in men. A possible mechanism is that testosterone levels decline with age and are inversely associated with leptin levels, thus leading to diminished food intake and an increased metabolic rate 41,42 .
Cooking and social engagement may affect both the quality and quantity of food intake 24 . We identified cooking to be a predictor of HANI only in women. Women usually prepare meals in Taiwan. The cooking involved requires at least planning, food choice, meal preparation itself, and various PAs 18,38 . Women deal with life alone more favourably than men 43 . The highest HANI participants and being more likely to live alone were women. By contrast, those registering the lowest HANI and living alone were men. Men tended to eat with others or consume fast food when they lived alone. Hence, eating with others seems crucial for men 44 . It may reduce the risk of malnutrition through social support, more food variety with improved dietary quality as well as QOL 45 . Altogether, cooking and eating with others may have combined benefits for nutritional status; these benefits are indicative of socio-psychological factors and therefore contribute to survival. The tangible support derived from eating with others has been associated with increased fruit and vegetable intakes in older men. Among women, emotional and informational support increased these intakes. By contrast, women have healthier diets when they cook for themselves 29 . This supports cooking as associated with more benefits for women and that eating with others is crucial for men.
We did not find the expected associations between HANI and cardiometabolic risk factors. Men with the highest HANI had higher triglycerides and diastolic blood pressure, and women with the highest HANI had higher total and LDL cholesterol. A lower cholesterol is associated not only with malnutrition, but also increased mortality in elders 46,47 . The advantage of HANI is that its predictive ability does not depend on cardiometabolic risk factors or detailed body compositional analysis 32 .
Our study has some strengths and limitations. First, we demonstrate that composite nutritionally modifiable factors, characterized by engagement and the ultimate consumption of a diversified diet, are conducive to higher survival. HANI, in this particular study context, may be a surrogate for associated health-promoting factors. Although we adjusted for several health domains, residual confounders likely remain, and HANI is unlikely to have a cause-and-effect relevance on its own. It must be modified if the community under investigation is socioculturally different. Second, it is both a limitation and a strength that HANI is gender-specific, a finding that demonstrates differences between older men and women. Third, people with a history of chronic diseases at baseline were not excluded. However, dietary habits that may change owing to disease have not been considered. It must be emphasized that the representative study population was free-living in the community, which means participants were functionally healthy. For policy relevance, community-based elderly have been the focus of our investigation. Nevertheless, the final indices did not change in the model adjusted for multi-morbidity (Charlson comorbidity index, Supplementary Table S2). Fourth, when we claim non-invasive assessment for the factors in HANI, we made anthropometric measurements to calculate BMI. Fifth, we examined the gender-specific indices with and without age for the development and validation sets and the entire study cohort. For women, after adjustment for age, the validation is less predictive of survival. This might be attributed to an over adjustment for age (twice: before and in the model). Another reason might be that the very old women cooked less, leading to a drop in their sample size and to statistical instability. Another consideration is that women outlive men and may represent an increasingly sociobiologically heterogeneous group with advancing years.
For senior citizens, HANI can offer a modifiable predictor of survival that is accessible, socioculturally adaptable, gender-specific, and may alter outcomes, although this would appear context-dependent. It comprises appetite, eating with others, dietary diversity score, and BMI for men, and cooking frequency, dietary diversity score, and BMI for women. The utility of HANI for the older population studied can be enhanced by the provision of an online assessment and monitoring tool. This tool can be used for several purposes such as nutritional education in the community or general population, and diagnosis of potential risk of nutritional disorder for further intervention in clinical settings. HANI is available for aged care policy makers and workers.

Methods
Cross-sectional and prospective study designs were used to evaluate the utility and validity of HANI. We recruited participants from the 1999-2000 Nutrition and Health Survey in Taiwan (NAHSIT). A total of 1937 older people aged 65 y or older completed face-to-face interviews. We randomly divided the participants by region into development (n = 966) and validation (n = 971) sets (Fig. 4) and underwent a physical examination in the morning. We collected their fasting blood for metabolic profiling. Informed consent was obtained from participants at interview. Ethical approval was obtained from the Institutional Review Boards of the National Health Research Institutes and Academic Sinica, Taiwan.
Nutritionally modifiable factors. We selected 11 nutritionally modifiable factors predicting mortality through prior knowledge of this population and a literature review. They were appetite 21 , chewing ability 22 , DDS 19 , daily vegetable expenditure 48 , frequency of cooking 18 , frequency of eating with others 44 , frequency of shopping 16 , alcohol consumption 49 , PA 10 , BMI 50 , and waist circumference 51 . The measurement details for 11 candidate factors are summarized in Table 1.
Dietary information was obtained through a one day 24-h dietary recall and a simplified food frequency questionnaire. Dietary quality was assessed using DDS based on a half serving of 6 food groups a day. The DDS ranged from 0 to 6, with a higher score representing better dietary quality 19 . We calculated participants' daily vegetable expenditure by using 24-h dietary recall 52 . Cut-off points for BMI and waist circumference were in accord with Taiwanese recommendations 53 . Participants with BMI (kg/m 2 ) <18.5 were considered underweight, 18.5-23.9 as normal weight, 24-26.9 as overweight, and ≥27 as obese. We defined normal waist circumference as waist circumferences of <90 and <80 cm in men and women, respectively. We compiled the measure of PA as metabolic equivalents (METs) per day. We classified the participants into 3 groups on the basis of daily METs: <1.5 (moderate PA < 30 min), 1.5-3 (moderate PA 30-60 min or vigorous PA < 30 min), and >3 (moderate PA ≥ 60 min or vigorous PA ≥ 30 min) 54 .
Other contextual factors. We measured health-related QOL by using the Chinese version of the Short Form (SF-36 ® ), modified for Taiwanese, which contains 36 self-assessment questions to measure 8 dimensions of health following the norm-based scoring system (μ = 50, σ = 10) 55 . A higher score indicates better QOL.
Cognitive impairment was assessed by a validated Short Portable Mental Status Questionnaire (SPMSQ) in Chinese. A total of 10 questions regarding orientation in time and place, personal history, long-term and short-term memory and calculation was use to evaluate mental health. Cognitive impairment was defined as ≥3 errors in the answers to the questions. Social engagement was assessed by 3 questions about visiting relatives, engaging in religious activities and involvement in social activities. 'Less social engagement' was defined as never being involved in these activities.   where height is measured in meters, resistance in Ohms, and age in years; men = 1 and women = 0. Resistance for whole body SMM was assessed by a BIA device (Parama-Tech BF-101) with two electrical signals (right wrist and right ankle). The equation was developed by Janssen et al. 56 and validated for Taiwanese elders by MRI-measured skeletal muscle mass 57

Healthy Ageing Nutrition Index (HANI).
We evaluated the association between the 11 candidate factors and all-cause mortality in the development set by using to step-by-step Cox proportional-hazard regression, in the following sequence (Fig. 4): 1. Candidate factor selection (refer to earlier text). 2. Age-adjusted hazard ratio: We determined the age-adjusted HR of each candidate factor. We selected the factors with age-adjusted HRs <1 for men and women. 3. Identified factors in a composite survival index: We entered these factors into a multivariable Cox proportional-hazard model manually to identify these factors according to their P values. 4. HANI scoring: We assigned the score of each selected factor by dividing each β-coefficient in the final model by the lowest β-coefficient and rounding to the nearest integer. We assigned HANI to each participant and summed the scores for all factors presented. 5. Cut-points: The cut-points for HANI was determined by the Youden index 58 .  . Continuous and categorical variables are expressed as means ± standard errors (SEs) and percentages, respectively. We evaluated the corresponding differences by ANOVA and chi-square test. Missing values in this study were principally due to survey design where two data collection activities, household questionnaire interview (n = 1937) and physical check-up (n = 2432), were combined and not always congruent. In order to achieve study power and not to over-estimate effects, we imputed missing data for candidate factors as the poorest group for categorical variables or mean for BMI and waist circumference in the same age by year and gender group. The distributions of each candidate factor with or without imputation in the development set by gender were not significantly different (Supplementary Table S5). The point estimates continued in the same direction after exclusion of participants with any missing HANI variable, significance disappeared, probably due to limited power (Supplementary Table S6). The follow-up time was from the date of interview to either the date of death or December 31, 2008. We assessed the association between HANI and all-cause mortality by Cox proportional-hazards regression model. Covariates adjusted were adjusted for age (in year), region, education level (illiterate, some up to primary school, and high school and above), current smoking status (yes, no), PA (<1.5, 2.5-2.9, ≥3 METs/d), cognitive impairment (yes, no), and less social engagement (yes, no). Because PA and ADL are highly correlated, we did not adjust for ADL in the models to avoid collinearity.
To evaluate the predictability of each component of HANI and the HANI score, a time-dependent receiver operating characteristic curve (ROC) analysis was used to discriminate between death and survivorship. This analysis uses sensitivity and specificity, both of which are time-dependent, to measure the predictability of a survival model as measured by the AUC 59 . We used C-statistics by logistic regression to evaluate the predictive accuracy and Harrell's C as well as Somers' D statistics for discriminatory performance (predictive power) of survival models 60 .
Data availability. The data that support the findings of this study are available from Academia Sinica and the Taiwan Department of Health, but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with the permission of Academia Sinica and Taiwan Department of Health.