Overall presentation Matram project

MATRAM projectModelisation and Evaluation

Bayesian Belief Network applied to Public Health Programmes

What tool for appraising the linkages between inputs/outputs and impact indicators ?

Overall picture

BBN: Probabilistic models based on directed graphs

-knowledge representation and comprehensive data analysis framework-visual structure attractive for exploring and explaining complex problems

Bayesia Lab : Powerful software for deep understanding of very complex and high-dimensional problem domains such as Public Health Programmes…

Sequence of the presentation

1- Three hats game

2- S&P Market shares

3- Perception of students on Family Medicine

4- Dataset from retired lung specialist

5- Malaria program

6- Holistic HIV program description

7- Summary

Common vocabulary

• Nodes represent the variables of the domain (e.g. the temperature of a device, a feature of an object, the occurrence of an event, the age of a patient)

• Links represent statistical (informational) or dependencies among the variables

The dependencies are quantified by conditional probabilitiesfor each node given its parents in the network.

1- Three hats game

Monthy Hall problem

1- Three hats game

1- Three hats game

1- Three hats game

P(Money Blue I Yellow hat empty) = P(Yellow hat to be selected I Money Blue ) X P (Money Blue ) / P(Yellow hat selected or empty)

P(Yellow hat to be selected I Money Blue ) = 1/2

P (Money Blue ) = 1/3 still

P(Yellow hat selected)

= 1/3 X ½

= 1/3 X 1 1/3 x ½ + 1/3 x 1 + 1/3 x 0 = 1/6 + 1/3 = 1/2 = 1/3 X 0

1

1

2a

2b

2c

2a

2b

2c

P(Money Blue I Yellow hat empty) =

(1/2 x 1/3) / 1/2 = 1/3

Because money must be somewhere P = 2/3 for RED !

1- Three hats game

1- Three hats game

-Conditional Probality distribution (child node) GIVEN THESE two nodes

Ability to update your knowledge given that you obtain new information …

2- S&P Market shares

Unsupervised learning : The goal of structure learning is to learn a model that best explains the data

Model built up by expertise: Complex and high-dimensional problem domain such as Public Health Programme

2- S&P Market shares

The wealth of data in the financial markets is a fertile ground for experimenting with knowledge discovery algorithms and for generating knowledge representations in the form of Bayesian networks

Database available: Daily closing prices of all stocks included in the S&P 500 index from

January 3, 2005 through December 30, 2010

2- S&P Market shares

Discritization by Kmeans

Maximum Spanning Tree

1st BayesiaLab User Conference

16/10/12

Strategic partnership : Procter & Gamble

16/10/12

3- Perception of students on Family Medicine

Database available: Database related to the questionnaires for the Perception

of Family Medicine study among Tajik students.

3- Perception of students on Family Medicine

2614 students interviewed

3- Perception of students on Family Medicine

Unsupervised Data Clustering

Creation of a model to represent uniform groups of students/observations

3- Perception of students on Family Medicine

All variables are discrete

Clustering process

3- Perception of students on Family Medicine

4- Dataset from retired lung specialist

Reasoning under uncertainty applies in two ways:• Diagnosis (inference from effect to cause)• Simulation (inference from cause to effect)

smoking has adirect influence onbronchitis

Model already set by expertise… right connectivity between pre-existing nodes

4- Dataset from retired lung specialist

5- Malaria program

Bayesian networks they can be

machine learned from data,

or,

can be built from human knowledge, i.e. from theory.

Cancer Care Ontario

16/10/12

Exploring New Methods of Identifying the Most Complex Patients for TargetedInterventions

Earlier identification of these patients could enable clinicians as well as policy makers to target clinical

or policy interventions aimed at reducing costlyhealth system use

Database: Clinical and past health system use data readily

available on the first day of dialysis

5- Malaria program

Practical example Source: Annuaire statistique Burkina Faso 2012

Impact indicator

5- Malaria program

5- Malaria program

Districts

Demographic/contextual parametersTotal Population

InfrastructureHealth CentersPrivate nursing based centers Private Medicines stores

Human ressourcesNurses with diploma% of norm compliance with expected HR for Health center

Malaria programmeInputTotal cases Palu Simp Cases Palu Comp Moins5 Cases Palu Comp Plus5OutputNbre TDR réalisés chez les <5ans Nbre de TDR réalisés chez les plus de 5ansNbre de GE réalisées chez <5ans Nbre de GE réalisées chez les plus de 5ansOutcomeNbre de TDR positifs chez les <5ans Nbre de GE positives chez <5ans Nbre deTDR positifs chez les plus de 5ans Nbre de GE positives chez les plus de 5ansNbre de cas traités avec ACT moins de 5ans dans les FSNbre de cas traités avec ACT plus de 5ans dans les FSImpactNbre cas Décès Palu Comp Moins 5 Nbre cas Décès Palu Comp Décès Plus 5

5- Malaria program

5- Malaria program

The assumed impact of ITNs on mortality is based on a systematic review of five randomly controlled

trials of ITNs in which child mortality was measured as an outcome.

Global Fund’s approach

to estimating lives saved…

The review concluded that ITNs contributed to a 17% reduction in “all-cause” mortality among children under

five years compared to a scenario where no nets were used at all.

Possibility to set up a holistic malaria program description…model

6- Holistic HIV program description

Global Fund KPI performance framework

The modelling exercise works by comparing the estimated effects of current service delivery scenarios on mortality

Against

a hypothetical “no treatment, no intervention” scenario

6- Holistic HIV program description

Until early 2011, the model assumed a standard survival rate of 86% at 12 months and 90% for each subsequent year for both adults and children on treatment.

Under the no-treatment scenario, it assumed a survival rate of 50% at 12 months and 0% at 24 months from the point of needing to start treatment.

6- Holistic HIV program description

The more context-specific approach will reveal the substantial variation in the effectiveness of HIV program.

This variation arises from probability distribution in :-access and uptake rates, -quality of care, -availability of skilled health workers, -rates of compliance with treatment, -lost to follow-up ,-population coverage of other determinants of clinical effectiveness, such as access to good nutrition

Variations, attribution (and causality): To what extend do changes in the inputs and processes of program & health

system result in changes in health outcomes ?

7- Summery

16/10/12

1- Mathematical model can be applied to Public Health programmes

Constraints:-Diversity and quality of the data-Representativeness of the Model (to be developped with Swiss TPH label)

2- Functionalities of BayesiaLAb do meet partner’s expectations in terms of impact evaluation

Constraints:-Mastery over the use of software-Tailored communication (against specific technical clients’ expectations)

Thank you for

your attention

What algorithms does BayesiaLab use for learning network parameters?

16/10/12

Mathematical background

What types of algorithms does BayesiaLab use for learning network structures?

BayesiaLab has developed a score-based learning algorithms. As opposed to the constraint-based algorithms that use independence tests to add or remove arcs between nodes, we utilize the MDL score (Minimum Description Length) to measure the quality of candidate networks with respect to the available data.This score, which is derived from Information Theory, allows to automatically take into account the data likelihood with respect to the network and the structural complexity of the network.

Mathematical background

All the optimization criteria of BayesiaLab’s learning algorithms are based on information theory (e.g. the Minimum Description Length). With that, no assumptions of linearity are made at any point.

16/10/12

Mathematical background

16/10/12

Mathematical background