Home About us Editorial board Search Ahead of print Current issue Archives Submit article Instructions Subscribe Contacts Login 
  • Users Online: 1792
  • Home
  • Print this page
  • Email this page

 Table of Contents  
Year : 2018  |  Volume : 4  |  Issue : 3  |  Page : 127-132

Delivering personalized dietary advice for health management and disease prevention

Department of Precision Medicine, RowAnalytics Ltd, Oxford, England, UK

Date of Web Publication18-Oct-2018

Correspondence Address:
Steve Gardner
RowAnalytics Ltd, C9 Glyme Court, Langford Lane, Kidlington, Oxford, OX5 1LQ
Login to access the Email id

Source of Support: None, Conflict of Interest: None

DOI: 10.4103/digm.digm_19_18

Rights and Permissions

Background and Objectives: Diet plays a huge role in health, both by increasing metabolic disease risks and acutely through adverse interactions with diseases and medications. Multimorbid and polypharmaceutical patients are at a particularly high risk of such interactions due to the number of drugs they take. This leads to avoidable hospitalizations and poor compliance. This study built and demonstrated a tool that provides personalized dietary advice that accounts for a patient's combination of disease and drugs in real-time on their mobile device. Methods: A comprehensive list of validated drug-disease-food interactions from several reputable sources was constructed. This was compiled into a knowledge graph using the RACE array logic platform. This interactions knowledge graph was used to power a personalized dietary advisor application on a mobile device. Results: Data from over 500,000 drug-disease-food interactions including 1,699 food ingredients and 9,526 disease interactions were compiled into a highly compressed knowledge model. This was used to inform recommendations for individual complex patients. It was also tested on virtual population of 10,000 multimorbid and polypharmaceutical patients. Conclusions: This study showed that digital health tools can provide highly contextual and adaptive responses from a single knowledge graph. The study showed it is possible to provide highly personalized health advice to complex patients in real-time on their own mobile device without having to hold such private information on a server. This enables highly secure, private and personalized digital health tools to be built.

Keywords: Adverse drug reactions, digital health, food, genomics

How to cite this article:
Gardner S, Pawlowski M, Møller GL, Jensen CE. Delivering personalized dietary advice for health management and disease prevention. Digit Med 2018;4:127-32

How to cite this URL:
Gardner S, Pawlowski M, Møller GL, Jensen CE. Delivering personalized dietary advice for health management and disease prevention. Digit Med [serial online] 2018 [cited 2023 Mar 24];4:127-32. Available from: http://www.digitmedicine.com/text.asp?2018/4/3/127/243635

  Introduction Top

It is difficult to overstate the importance of diet and lifestyle on health or the challenges that long-term poor choices and behaviors can create for public health and social care systems. As populations have become older and more affluent, they tend to adopt more sedentary lifestyles and less healthy diets. This has led to predictable increases in several chronic diseases, which now impose a huge burden on health and social care systems.[1] Preventable lifestyle-induced diseases with major dietary components, such as cardiovascular disease, Type 2 diabetes, cancer, and dementia now account for 65% of all health-care visits and 90% of deaths in the UK.[2] About 86% of all healthcare spending in the US covers people with chronic diseases, costing over $750B/year to treat.[3] Chronic noncommunicable diseases are becoming major health issues in developing economies.[4]

Patients' diseases, drugs, foods, and microbiome interact and influence one another profoundly.[5] Chronically, these factors and their interactions modulate disease risks for some of the most costly and prevalent metabolic diseases through the activation of processes such as autoimmune responses, inflammation, and insulin resistance.[6] Other effects may be mediated by changes to the gut microbiome, affecting absorption of key nutrients or the production of onco- and neuro-protective factors such as short-chain fatty acids. Butyrate, a product of gut microbiota, has, for example, been demonstrated to be oncoprotective for colorectal cancer[7] and also implicated in protection against dementia.[8] The shift in Eastern diets over the last 30 years away from rice and other plant-based foods, has led to reduced levels of complex starches in the lower bowel[9] and dysbiosis of the microbiota that produce the protective butyrate. In the same period, we have seen sharp rises in the incidence of colorectal cancer in these populations.[10]

More acutely, drug, disease, and food interactions can cause serious adverse events leading to expensive emergency hospitalizations and/or death.[11] Approximately 65% of hospital admissions of elderly patients are drug-related, with 11% attributed directly to adverse drug reactions (ADRs).[12] These admissions account for between 3% and 16% of hospitalizations[13],[14] and are estimated to cost the US healthcare system over $100B/year.[15] Epidemiological and meta-analysis studies suggest that at least 50% of these admissions are preventable.[16]

It is extremely challenging for clinicians to predict all the potential interactions in the context of a specific patient. In part, this is due to a lack of tools and time during consultations, and also due to a lack of training - nutrition and dietetics is almost wholly absent from the standard training syllabus for clinicians. Mainly the challenge stems from the large number and complexity of interactions and the huge number of confounding factors. The result is that very few patients ever receive personalized dietary advice as an integrated part of their treatment regimen or health management plan. Standard public health messages such as the UK's “eat 5 (portions of fruit and vegetables) a day” do not account for and may even conflict with the personal requirements of polypharmaceutical patients.

Elderly and multimorbid adults are particularly vulnerable to ADRs due to the number of drugs they are prescribed[17] and over-the-counter and herbal medications and supplements they are taking. Identifying these potentially harmful drug-disease-food interactions is not as simple as looking up a list of all known interactions. Even ostensibly simple and well-known interactions – for example “green leafy vegetables reduce the efficacy of warfarin” – are complex. Warfarin interferes with the liver's ability to process vitamin K (which occurs in high levels in leafy greens) and inhibits production of blood clotting proteins. High dietary levels of vitamin K overcome the inhibitory effect of warfarin, leading to more natural clotting and leaving some patients exposed to increased the risk of stroke.[18] Multiple other factors such as painkillers, antibiotics, dietary supplements, or alcohol also affect the clotting cascade and vitamin K levels. Vitamin K itself also goes on to affect other processes in the body and impacts on liver disease, cystic fibrosis, diarrhea, and digestive conditions,[19] which in turn affect how foods are absorbed. The network of connections is therefore intricately intertwined and difficult to deconvolute and test comprehensively.

It is also well-known that generic lifestyle and dietary advice is not just ineffective at changing behavior but may be actively ignored or mistrusted by patients.[20] When provided, it is often insufficiently engaging, accessible and actionable and tends to have a generic wellness or single disease (e.g., diabetes or obesity) focus, rather than considering the patient's personal diseases, medicines, and risk factors. This fails to effectively communicate risks and personalize dietary advice in a form that people can remember, engage with and use in their daily lives. A tool that can provide real-time personalized advice accounting for a full range of the patient's circumstances is therefore required.

The precisionlife DIET system was developed to provide patients and carers with this personalized advice on demand as to whether particular foods or ingredients are going to present a risk of interactions for them and to suggest alternative products that will not cause such an issue for them. This on-going advice aims to educate patients in the management of their disease/s and engage them as a critical partner in maintaining their health.

  Methods Top

Providing this level of personalized dietary advice is complex and requires some level of access to information describing patient's drug and diseases. To better protect, the patient's private medical data, it should ideally never leave their possession, i.e., the storage of their encrypted profile and all of the associated computation should be fully managed on their own device, without using a web service or other server. Due to the complexity of the interactions being computed and the need to personalize the advice, this requires a very efficient computational engine to perform the complex decision support.

The RACE constraint engine,[21] which is based on an array logic approach, was used as it solves complex constraint problems on finite domains very efficiently. This approach scales near linearly with increasing dimensionality, deducing all logical information very efficiently by simple parallel array operations rather than greedy heuristic searches, which would scale exponentially.

The constraint engine first compiles the complete solution space, testing it for logical consistency, and identifies any invalid information or unreachable outcomes. It then creates a highly compressed look-up table of all valid solutions using a nested array representation. Runtime operation of the RACE Application Programming Interface (API) then only requires simple indexed lookups, giving an instant response and a small memory footprint even on a mobile device or smartphone.

To generate the knowledge model, a set of scientifically validated sources of drug-disease-food interactions was identified and descriptions of the specific interactions mined. Information describing all known serious or moderate interactions between drugs and other drugs, diseases and foods is published by regulatory agencies such as FDA, EMEA, and MHRA. Over 10 of these scientifically validated data sources from different countries were integrated and semantically normalized to identify potential drug-disease-food interactions. In addition, where known, their basic mechanisms of action and mapping to organs were documented, so that potential for cumulative inhibitory and excitatory interactions can be inferred. In total information on over 8,500 drug formulations (including generic drugs and prescription/trade names) was included in this study.

The potential interactions between all possible multiple combinations of foods, drugs, and diseases based on this set of known interactions were then compiled into a knowledge model. This compilation process eliminates all combinations that generate invalid solutions according to the constraints applied (i.e., the reported interactions). The logic of the model was inverted so that the resulting solution space contained only the combinations that are known to cause problems, as there are many fewer problematic combinations than nonproblematic ones. As shown in [Figure 1], this generated a highly compact (428Kb) nested array knowledge model which provided the analytical and predictive substrate to power the precisionlife DIET decision support app on the mobile device.
Figure 1: Process for compiling the precisionlife DIET nested array model containing all drug-disease-food and other interactions

Click here to view

A large database of food products and recipes, including ingredients and nutritional information, was generated by mining a variety of basic and processed foods from several food product datasets including retailers, brands, food product listings, and nutrition sites. The nonredundant database contains around 185,000 basic and branded food items predominantly sold in the UK and US markets. The ingredient and allergen composition of each of the food items was either mined from food composition databases or inferred from labels and/or the nearest similar product (identified using a semantic match algorithm). Food products were divided into six major categories consistent with the structure of the Tesco Groceries website (Fresh Food, Frozen Food, Bakery, Food Cupboard, Drinks, and Baby Food Products). Each of these categories were divided into three additional layers of sub-categories to give 464 detailed subcategories. Example food classifications include “Drinks/Alcoholic/Wine/Red Wine” and “Fresh Food/Fresh Meat/Fresh Beef/Minced Beef.”

To provide a mix of accuracy and flexibility, food products were matched against the food product database using a test cascade of barcodes, EAN-13 codes, branded product names, exact product name matches, or nearest semantic matches. In this semantic matching, each of the food products is described as a 300-dimensional vector (where a feature may be the department, aisle, shelf, the ingredients and/or their component 1, 2, or 3 word n-grams, as well as words from the product name, etc). Products sharing similar vectors using a cosine distance metric are predicted to be close matches. For example, [Table 1] shows the top matches to a previously unknown alcohol-free drink product are all alcohol-free variants of usually alcoholic drinks.
Table 1: The closest semantic matches to an unknown alcohol-free drink product

Click here to view

The user's personal profile of drugs, diseases, and allergies as entered into the mobile app was used as an input state vector that was applied to the knowledge model by the RACE API. All potential interactions between a specified set of food products and their ingredients (for example in an online shopping basket or identified through a barcode scan) and the patient's combination of diseases and drugs were identified in real-time (taking <20 ms on a mobile phone) as shown in [Figure 2].
Figure 2: Process of generating recommendations for one or more selected food products by checking them for interaction potential and severity using the RACE runtime API to apply the user's profile as an input state vector to the interactions nested array model

Click here to view

If the food product has the potential to cause an interaction, personalized dietary advice can be offered to the user in the form of red/amber/green alerts and customizable “swaps” as shown in [Figure 3]. These swaps are alternative food products that are from the same product category (or supermarket shelf) but which do not share the potential risk for the drug: food or disease: food interactions of the originally chosen item for the user. The user can be rewarded for choosing such alternative healthier products as a way of gamifying the tool.
Figure 3: (a) Highlighted foods showing the interaction potential with a user's profile categorized from high risk/severity (red), medium risk/severity (amber) and no risk (green). (b) Suggested food swaps for the alcoholic drink product

Click here to view

Given the number and complexities of drug-disease-food interactions, there is often no product in a food subcategory that will be safe for everyone. To avoid suggesting products that will present a similar interaction risk for the user, for example, substituting kale for broccoli for a patient on warfarin, or smoked bacon for salami for a patient on isocarboxazid, every potential swap is tested against the user's profile. This makes sure that all suggestions are appropriate alternatives for the user.

  Results Top

In this study over 500,000 drug-drug interactions and 136,784 drug-disease interactions were identified. When converted to a nonredundant database based on the drugs' Active Pharmaceutical Ingredients (e.g., acetylsalicylic acid is sold in Bayer Aspirin, Bayer Children's Aspirin, Bufferin, Easprin, and Ecotrin) this was reduced to 9,526 drug-disease interactions representing interactions of 2,545 drugs' active ingredients and 1,699 food ingredients. All of these were incorporated into the knowledge model. This was implemented as a secure mobile phone application providing real-time advice based on a user's profile.

The resulting drug-disease-food interactions knowledge model and recommendations were reviewed initially using automated quality assurance protocols and then tested extensively on a virtual population of 10,000 complex (multi-morbid and polypharmaceutical) patients with a randomized but accurately distributed set of morbidities. This population was constructed with the UK normal distribution of sex, age, diseases (with stage and severity) and prescriptions (following NICE standard clinical pathway guidelines for their assigned diseases).

precisionlife DIET accurately identified all the “obvious” first-order reported drug: food interactions, for example, grapefruit for simvastatin users, as well as most for some foods with obfuscated or ambiguous ingredients. It also provided accurate warnings in a number of cases where the interactions of multiple food ingredients, drugs and/or disease processes could result in a problem for the user, for example, hypocalcemia for patients with chronic kidney disease on certain diuretics.

It was necessary to implement a refinement to restrict some of the interaction warnings based on the amount of a particular ingredient in foods. For example, wine-based vinegars are commonly used as flavorings, but the proportion of alcohol is small (<2% by volume) and the volume used is typically very small (<10 mL). This poses a negligible risk in most cases even where alcohol is contraindicated. The issue was overcome by restricting the recommendations based only on ingredients in the first five positions on the label (which are ordered by volume).

  Discussion Top

The same engine can be used for wellness applications to maintain a healthy diet. This may include nutritional guidelines with specific targets for nutritional components. It may also be used to enable people to achieve certain health goals, for example, the incorporation of additional folic acid into the diet of a newly pregnant woman or management of GI and total carbohydrates in a prediabetic person. Information on 457 allergen labels and 19,612 pharmacogenetic interactions is also being included in more recent versions (e.g., for future use with consumer or clinical genomics profiles once they become available).

  Conclusions Top

The study has shown that it is possible to provide accurate personalized dietary advice to complex patients wholly on their own mobile device in real-time and without needing to maintain or even see their personal medical data.

Future versions could enable social and clinical care providers to quickly optimize menu selections across all of their beds, taking into account each patient's nutritional requirements and the need to avoid their personal drug-disease-food interactions. It could help improve patient outcomes, reducing recovery times through better nutrition and lower incidence of food interactions and allergies.[22] Considerations such as swallowing difficulties (dysphagia) and altered taste/smell perception could also be accommodated on a patient by patient basis by including textual and sensory selection criteria within the overall menu plan.

Financial support and sponsorship

This project was supported in part by research funding from the European Commission the FI-ADOPT programme.

Conflicts of interest

All authors are employees of RowAnalytics Ltd which developed the precisionlife DIET system.

  References Top

Prince MJ, Wu F, Guo Y, Gutierrez Robledo LM, O'Donnell M, Sullivan R, et al. The burden of disease in older people and implications for health policy and practice. Lancet 2015;385:549-62.  Back to cited text no. 1
World Health Organization. Non-Communicable Diseases Progress Monitor2017. Geneva: WHO; 2017. Available from: http://www.who.int/nmh/publications/ncd-progress-monitor-2017/en/ [Last accessed on 2018 Aug 08].  Back to cited text no. 2
Gerteis J, Izrael D, Deitz D, LeRoy L, Ricciardi R, Miller T, Basu J. Multiple Chronic Conditions Chartbook. AHRQ Publications No, Q14-0038. Rockville, MD: Agency for Healthcare Research and Quality; 2014. Available from: https://www.ahrq.gov/sites/default/files/wysiwyg/professionals/prevention-chronic-care/decision/mcc/mccchartbook.pdf. [Last accessed on 2018 Aug 08].  Back to cited text no. 3
Pradeepa R, Mohan V. Prevalence of type 2 diabetes and its complications in India and economic costs to the nation. Eur J Clin Nutr 2017;71:816-24.  Back to cited text no. 4
Le Bastard Q, Al-Ghalith GA, Grégoire M, Chapelet G, Javaudin F, Dailly E, et al. Systematic review: Human gut dysbiosis induced by non-antibiotic prescription medications. Aliment Pharmacol Ther 2018;47:332-45.  Back to cited text no. 5
Esser N, Legrand-Poels S, Piette J, Scheen AJ, Paquot N. Inflammation as a link between obesity, metabolic syndrome and type 2 diabetes. Diabetes Res Clin Pract 2014;105:141-50.  Back to cited text no. 6
Wu X, Wu Y, He L, Wu L, Wang X, Liu Z. Effects of the intestinal microbial metabolite butyrate on the development of colorectal cancer. J Cancer 2018;9:2510-7.  Back to cited text no. 7
Liu J, Sun J, Wang F, Yu X, Ling Z, Li H, et al. Neuroprotective effects of clostridium butyricum against vascular dementia in mice via metabolic butyrate. Biomed Res Int 2015;2015:412946.  Back to cited text no. 8
Zhai FY, Du SF, Wang ZH, Zhang JG, Du WW, Popkin BM, et al. Dynamics of the chinese diet and the role of urbanicity, 1991-2011. Obes Rev 2014;15 Suppl 1:16-26.  Back to cited text no. 9
Fang JY, Dong HL, Sang XJ, Xie B, Wu KS, Du PL, et al. Colorectal cancer mortality characteristics and predictions in China, 1991-2011. Asian Pac J Cancer Prev 2015;16:7991-5.  Back to cited text no. 10
Patel TK, Patel PB. Mortality among patients due to adverse drug reactions that lead to hospitalization: A meta-analysis. Eur J Clin Pharmacol 2018;74:819-32.  Back to cited text no. 11
Alhawassi TM, Krass I, Bajorek BV, Pont LG. A systematic review of the prevalence and risk factors for adverse drug reactions in the elderly in the acute care setting. Clin Interv Aging 2014;9:2079-86.  Back to cited text no. 12
Bednall R, McRobbie D, Hicks A. Identification of medication-related attendances at an A & E department. J Clin Pharm Ther 2003;28:41-5.  Back to cited text no. 13
Nelson KM, Talbert RL. Drug-related hospital admissions. Pharmacotherapy 1996;16:701-7.  Back to cited text no. 14
Formica D, Sultana J, Cutroneo PM, Lucchesi S, Angelica R, Crisafulli S, et al. The economic burden of preventable adverse drug reactions: A systematic review of observational studies. Expert Opin Drug Saf 2018;17:681-95.  Back to cited text no. 15
Hakkarainen KM, Hedna K, Petzold M, Hägg S. Percentage of patients with preventable adverse drug reactions and preventability of adverse drug reactions – A meta-analysis. PLoS One 2012;7:e33236.  Back to cited text no. 16
Oscanoa TJ, Lizaraso F, Carvajal A. Hospital admissions due to adverse drug reactions in the elderly. A meta-analysis. Eur J Clin Pharmacol 2017;73:759-70.  Back to cited text no. 17
Chang CH, Wang YW, Yeh Liu PY, Kao Yang YH. A practical approach to minimize the interaction of dietary Vitamin K with warfarin. J Clin Pharm Ther 2014;39:56-60.  Back to cited text no. 18
Krzyżanowska P, Drzymała-Czyż S, Pogorzelski A, DuśŻuchowska M, Skorupa W, Bober L, et al. Vitamin K status in cystic fibrosis patients with liver cirrhosis. Dig Liver Dis 2017;49:672-5.  Back to cited text no. 19
Mozaffarian D, Angell SY, Lang T, Rivera JA. Role of government policy in nutrition-barriers to and opportunities for healthier eating. BMJ 2018;361:k2426.  Back to cited text no. 20
Jensen CE, Møller GL. European Patent Number EP1062603B9. European Patent Office; 1998.  Back to cited text no. 21
Martin L, Gillis C, Atkins M, Gillam M, Sheppard C, Buhler S, et al. Implementation of an enhanced recovery after surgery program can change nutrition care practice: A multicenter experience in elective colorectal surgery. JPEN J Parenter Enteral Nutr 2018. doi: 10.1002/jpen.1417. [Epub ahead of print].  Back to cited text no. 22


  [Figure 1], [Figure 2], [Figure 3]

  [Table 1]


Similar in PUBMED
   Search Pubmed for
   Search in Google Scholar for
 Related articles
Access Statistics
Email Alert *
Add to My List *
* Registration required (free)

  In this article
Article Figures
Article Tables

 Article Access Statistics
    PDF Downloaded212    
    Comments [Add]    

Recommend this journal