A Bayesian network model to segment consumers
based on their minor digestive concerns
Pierrick Rivière1, Fabien Craignou2, Peter Whorwell3, Helen Carruthers3
1 Sensory and Behavior Science, Danone Research, RD 128, 91767 Palaiseau – France
2 Repères, 9 rue Rougemont, 75009 Paris – France
3 Department of Medecine, University Hospital of South Manchester, Manchester, M23 9LT UK
context & objectives
Beyond sensory pleasure and nutritional intake, food can also integrate health functionalities like improving digestive health. This expectation is not negligible : 60% of English
women claim to suffer from minor digestive disorders! Most of them do not consult health care professionals but try alternative solutions, including food. Designing a product to
match these expectations requires an accurate knowledge of the troubles experienced by consumers, to complement the existing medical knowledge. Minor digestive troubles
cannot be accessed in a straight-forward manner as naïve consumers don’t have the medical knowledge nor the ability to describe/name their digestive troubles[1].
Nevertheless, consumers can translate what they feel when experiencing digestive troubles.
The objective is to build a precise typology of minor digestive troubles based on perceived & experienced sensations described by consumers.
methodology
1000 English women nationally representative declaring Experienced sensations have been structured into
“tell us the story of your last digestive concern”…
minor digestive trouble. attributes based on following requirements :
SYMPTOM DESCRIPTION - one-dimensional (a single sensation)
On-line interactive survey - accurate / non ambiguous
• Experienced sensations (40 items list)
- combining imagery & concrete description
Focus on the last trouble to increase accuracy of the • Digestive concern illustration (12 pictures) - based on layman words
description. • Etiology (supposed causes) - uses the most frequent words / expressions
• Impact (Emotional & social consequences; >> examples :
• Questionnaire items related to the personal frequency / pain) “I burp”
experiences of digestive troubles extracted from a • Context “I have gurgling. It's turning inside me, as if nothing is in the right
previous Qualitative Survey SOLUTIONS & EXPECTATIONS place”
…
USER PROFILING (demographics; IBS detection)
data analysis & results
1-Identifying groups of experienced sensations likely to happen together 2-Identifying groups of consumers
= The symptoms concerned by the same combinations of symptoms
Bayesian Clustering of Consumers (BayesiaLab Data Clustering)
Automatic Bayesian Learning (BayesiaLab EQ algorithm)
[2] Latent Class Model, in which consumer clusters are connected to the symptoms.
Discovering probabilistic relations between symptoms
NO a-priori relation to be defined: focus on what consumers really feel.
EM algorithm to determine the parameters of the model.
•Conservative model[3] : compromise between the data fit and the structural complexity
Pseudo-random walk between 2 and 20 to determine the number of clusters.
•10-fold cross-validation to ensure the robustness of the structure.
Stomach rock
Stomach rock Swollen tummy
Swollen tummy
6%
hard
hard
Cluster 12 10%
IIcan’t evacuate my bowels I’m constipated
My belly is hard
My belly is hard Skin of my tummy is tensed
IIhave a full stomach
have a full stomach can’t evacuate my bowels I’m constipated Cluster 11
Skin of my tummy is tensed
Quite painful to touch stomach
9%
Quite painful to touch stomach IIfeel full Emptying my bowels is difficult
feel full Emptying my bowels is difficult Cluster 10 10%
Knots in my intestine
Knots in my intestine Tummy is going to explode
Tummy is going to explode
Can cause haemorrhoids Evacuation is painful
Can cause haemorrhoids Evacuation is painful
Cluster 9 6% Cluster 6
Intestine is in spasm
Intestine is in spasm
7%
Stagnation within my tummy
Stagnation within my tummy
Need to relieve the pressure
Need to relieve the pressure
Spasms inside me
Spasms inside me There is food in my stool, not fully
There is food in my stool, not fully 5% Cluster 7
Stomach is blowing up like a balloon
digested
digested
Cluster 8
Stomach cramps
Stomach cramps
Stomach is blowing up like a balloon 7%
As if my stool has fermented inside me
As if my stool has fermented inside me
It stretches my tummy, like contractions
It stretches my tummy, like contractions
Cluster 5
IIhave gurgling
have gurgling Food doesn’t go through me
Food doesn’t go through me 7% 12%
IIget a pricking inside me
get a pricking inside me
Cluster 3
Small hair bubbles inside of me
Small hair bubbles inside of me
It is burning inside my tummy
It is burning inside my tummy
Going to the toilet is almost
Going to the toilet is almost
explosive
Cluster 4
explosive Colour variation:
11% purity of the cluster
Acid in me so that’s burning IIhave often smelly wind
have often smelly wind Really urgent need to go to the toilet
Really urgent need to go to the toilet
Acid in me so that’s burning (the darker, the purer)
IIburp
burp IIfeel full of wind
feel full of wind IIoften have gas
often have gas
Large amount of stool
Large amount of stool Cluster 2 10% Positioning:
I’ve got acid reflux
I’ve got acid reflux probabilistic proximity
Body feels really heavy
Body feels really heavy
Regurgitation in the throat
Regurgitation in the throat
IIcannot help but pass wind
cannot help but pass wind Cluster 1
11 groups of symptoms were identified. 12 groups of consumers emerged,
each concerned with a specific symptoms profile = a digestive trouble
Each group can be summarized with a latent variable, Each group of consumers can be interpreted with a probabilistic symptom profile,
which can be interpreted in a probabilistic way. for an easy and compact communication. Each cluster is characterized by a
specific & contrasted combination of perceived symptoms.
ACID REFLUX
33% of consumers are Yes
EXAMPLE Symptom did not occur 67.0% concerned with acid reflux
symptoms (a priori probability) ACID REFLUX
Symptom occurred 33.0%
+ +
+ + +: significant
If for example, consumer X answers… difference
I’ve got acid reflux -----------------No Yes against a priori
Probability that probability
Acid in me so that's burning ------Yes consumer X is affected
Regurgitation in the throat---------Yes by acid refluxis 81%.
I burp -------------------------------No GURGLING
+
+ +
C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12
conclusions OTHER SYMPTOMS …
This approach provides an accurate & useful
description of minor digestive concerns in the UK pop.
The 12 clusters are fully interpretable by
Bayesian networks allowed identifying
digestive troubles based on symptom
outlooks
gastroenterologists & physiologically relevant. combinations without any a priori. Despite the gap between the medical vision and the layman
representation, this approach allows the connection of consumer
They are used by R&D teams to translate consumer Probabilistic reasoning and graphical perceptions to the expert physiological knowledge.
needs into new researches & clinical studies. models help communication.
This link provides a complementary understanding of the human body
[1] R. Monrozier, A. Bonnet, I. Boutrolle, N. Boireau. (2009).Toward a consumer typology of health concerns. An application to minor digestive disorders. Poster - 8th Pangborn Sensory Symposium
[2] J. Pearl and S. Russel, 2000 "Bayesian networks" , UCLA Cognitive Systems Laboratory, Technical Report (R-277), November 2000. In M.A. Arbib (Ed.), Handbook of Brain Theory and Neural Networks, Cambridge, MA: MIT Press, 157-160, 2003.
[3] Friedman N., Goldszmidt M., “Learning Bayesian networks with local structure ”, Proc. of the 12th Conf. on Uncertainty in Artificial, Morgan Kaufmann, 1996.