Translate Register Log In Login with facebookLogin and Register

Vitamin D estimation nicely improved by neural networks – May 2015

Derivation and Validation of a Clinical Diagnostic Tool for the Identification of Older Community-Dwellers With Hypovitaminosis D

J Am Med Dir Assoc. 2015 Apr 23. pii: S1525-8610(15)00222-4. doi: 10.1016/j.jamda.2015.03.008. [Epub ahead of print]
by The Society for Post-Acute and Long-Term Care Medicine.
Cedric Annweiler MD, PhDa,b’ CeAnnweiler at chu-angers.fr , Anastasiia Kabeshova MSa, Mathilde Legeay MDc, Bruno Fantino MD, PhD a, Olivier Beauchet MD, PhD a
a Division of Geriatric Medicine and Memory Clinic, Department of Neuroscience, UPRES EA 4638, UNAM, Angers University Hospital, Angers, France
b Department of Medical Biophysics, Robarts Research Institute, Schulich School of Medicine and Dentistry, The University of Western Ontario, London, Ontario, Canada
c School of Medicine, University of Angers, Angers, France

Objectives: Hypovitaminosis D is highly prevalent among seniors. Although evidence is insufficient to recommend routine vitamin D screening in seniors, universal vitamin D supplementation is not desirable either. To rationalize vitamin D determination, our objective was to elaborate and test a clinical diagnostic tool for the identification of seniors with hypovitaminosis D without using a blood test.

Design: Derivation of a clinical diagnostic tool using artificial neural networks (multilayer perceptron; MLP) in randomized training subgroup of Prevention des Chutes, Reseau 4' cohort, and validation in randomized testing subgroup.

Setting: Health Examination Centers of health insurance, Lyon, France.

Participants: A total of 1924 community-dwellers aged >65 years without vitamin D supplements, consecutively recruited between 2009 and 2012.

Measurements: Hypovitaminosis D defined as serum 25-hydroxyvitamin (25OHD) concentration < 75 nmol/L, <50 nmol/L, or <25 nmol/L. A set of clinical variables (age, gender, living alone, individual deprivation, body mass index, undernutrition, polymorbidity, number of drugs used daily, psychoactive drugs, biphosphonates, strontium, calcium supplements, falls, fear of falling, vertebral fractures, Timed Up and Go, walking aids, lower-limb proprioception, handgrip strength, visual acuity, wearing glasses, cognitive disorders, sad mood) were recorded. Several MLPs, based on varying amounts of variables according to their relative importance, were tested consecutively.

Results: A total of 1729 participants (89.9%) had 25OHD <75 nmol/L, 1288 (66.9%) had 25OHD <50 nmol/L, and 525 (27.2%) had 25OHD <25 nmol/L. MLP using 16 clinical variables was able to diagnose hypovitaminosis D < 75 nmol/L with accuracy = 96.3%, area under curve (AUC) = 0.938, and k = 79.3 indicating almost perfect agreement. It was also able to diagnose hypovitaminosis D < 50 nmol/L with accuracy = 81.5, AUC = 0.867, and k = 57.8 (moderate agreement); and hypovitaminosis D < 25 nmol/L with accuracy = 82.5, AUC = 0.385, and k = 55.0 (moderate agreement).

Conclusions: We elaborated an algorithm able to identify, from 16 clinical variables, seniors with hypovitaminosis D.


VitaminDWiki purchased a copy of the PDF
Study focused on a subset of the entire population: seniors living in community setting
They ignored seniors who were already taking vitamin D

without NN
NN accuracy
< 75 nmol90%96%
<50 nmol67%82%
<25 nmol27%82%

The important variables included:
Being in a union, Gender, Age, Height, EPICES score, Number of drugs taken per day
Time to perform the TUG, Use walking aids, Abnormal Clock Drawing Test, Sad mood
Feeling of emptiness, Happiness, Hopelessness, 4-item Geriatric Depression Scale score
Fear of falling, History of vertebral fractures, BMI, Living alone, Handgrip strength
Weakness, Mean visual acuity

Many other variables were not used by the neural network,

VitaminDWiki is considering making a system of neural networks 

Would integrate data from dozens of existing data sets - which use very different data types and definitions
The resulting system could be used by anyone visiting VitaminDWiki
    to estimate current level and suggest dose needed to get to 40 nanograms
The more data the visitor would enter, the more accurate the results would be
Accuracy would be greatly improved if test results are entered
It may include recomendations for cofactors, loading dose, etc. based on actual data.

VitaminDWiki articles which hint to other variables to be considered

See any problem with this page? Report it (FINALLY WORKS)