Modeling Malaria: Methods and Applications

class: center, middle, inverse, title-slide

# Modeling Malaria: Methods and Applications
## SNRE Exit Seminar<br/> (Slides available at: bit.ly/bentoh-exit)
### Kok Ben Toh
### 2020/09/21

---

class: inverse, center, middle

# Introduction to Malaria

---
# Introduction to Malaria

.pull-left1[
- Caused by protozoan parasite *Plasmodium*

- 5 species that causes human malaria

- Mosquito-borne disease
]

.pull-right1[
<img src="../assets/img/malaria-life-cycle.jpg" width="1065" style="display: block; margin: auto;" />
.source[*[Johns Hopkins Bloomberg School of Public Health](http://ocw.jhsph.edu)*]
]

---
# Stats and Geography

- 230 million cases worldwide in 2018; 400k death, 2/3 under 5 years old

- Distribution of disease
  - Tropics: Sub-Saharan Africa, South and Southeast Asia, Latin America
  - Correlated to environmental and socio-economic factors
  
<img src="../assets/img/world-malaria-map.png" width="95%" style="display: block; margin: auto;" />
.source[*(Weiss et al. 2019)*]

---
# Modeling Malaria

.pull-left[
- Integral part of understanding malaria

- Mathematical, e.g.
  - Compartmental deterministic model to understand disease transmissions

- Statistics
  - Risk factor analysis
  - Mapping malaria risks
  - Clinical diagnosis
]

.pull-right[
<img src="../assets/img/malaria-compartment.jpg" width="1600" style="display: block; margin: auto;" />
.source[*(Mandal et al. 2011)*]
]

---
# Overview

- Improving National Level Spatial Mapping of Malaria through Alternative Spatial and Spatio-temporal Models

- Interlude

- Guiding placement of health facilities using malaria criteria and interactive tool

- Decision Support Tool to Predict Causes of Childhood Febrile Illness Using a Bayesian Model Approach

---
class: inverse, center, middle

# Improving National Level Spatial Mapping of Malaria through Alternative Spatial and Spatio-temporal Models

---
# Mapping malaria

- Malaria risk map guides efficient resource allocation for intervention program

---
# Choice of Model

- Most common approach: Gaussian Process Model
  - Points close to the observed data point should share similar prevalence

- Time consuming as sampling point large
  - Commonly approximated using Stochastic Partial Differential Equation (SPDE) + Integrated Nested Laplace Approximation (INLA)

---
# Alternative model?

.pull-left[
<img src="../assets/img/spde-gamgbm-length.png" width="50%" style="display: block; margin: auto;" />
]

.pull-right[
- SPDE-INLA is complicated

- Can we use faster and simpler alternatives instead?
]

---
# Alternative model 1

- Spatial smoothing using Generalized Additive Model (GAM)
  - An underlying smoothed "landscape" unexplained by the covariates

---
# Alternative model 2: Machine learning

- Decision-trees ensemble methods
  - E.g. Random forests, gradient boosting tree/method (GBM)

---
# Borrow strength from the past?

- More and more countries now have multiple years of national malaria survey data.
  - e.g. Burkina Faso 2010 vs 2014

- Spatial vs Spatiotemporal setting

---
# Objectives

- To determine if GAM and state-of-the-art machine learning method (e.g. GBM), under both spatial and spatio-temporal setting, can be good alternatives to the more complicated SPDE method

- To determine if inclusion of past data is beneficial in modeling the current spatial distribution of malaria prevalence at the national scale

---
# Data

- Demographic Health Survey data from five countries

- Spatial covariates: free and publicly available remote sensing or GIS products

.pull-left[
<img src="../assets/img/map-chp1-5cty.png" width="1365" style="display: block; margin: auto;" />
]

.pull-right[
<img src="../assets/img/chp1-covars-layer.png" width="1365" style="display: block; margin: auto;" />
]

---

# Model comparison

- 5 countries × 4 models × 2 settings
- Spatial setting

- Spatiotemporal setting

---
# Results

- GAM and SPDE 👍

---
# Results

- GAM and SPDE 🙌

---
# Results

- Very small difference among top models (look at the axis)

---
# Results

- GBM 😲

---
# Results

- GBM 😲 GAM 😨

---
# Spatial vs Spatiotemporal setting

---
# Discussions

- No single best model, performance varies from setting to setting and country to country

- SPDE is consistent, but doesn't gain much from incorporating past data
  - Can deteriorate when past and present spatial dependency are very different
  
- GAM is good, fast and simple-to-use alternative, especially with more data
  - Dismal performance in irregularly shaped countries
  - High perimeter to root area ratio: Malawi 8.7, Uganda 5.7

- GBM unpredictable but generally fits well with more data available

- 🏠💬 Fit multiple model or at least benchmark with GAM

---
class: inverse, center, middle

# Interlude

---
# Malaria treatment

- The most famous antimalarial drug is... Chloroquine

- But currently, the state-of-the-art drug is the *Artemisinin (Qinghaosu, 青蒿素)* and derivatives.

- First isolated in 1972 by Tu Youyou (屠呦呦) and members of *Project 523*

- Awarded 2015 Nobel Prize in Medicine

.source[*Internet, creator unknown*]

---
# Artemisinin

- Isolated from *Artemisia annua*

- Traditional Chinese Medicine

- Prescription for malaria in 4th century AD

.source[*[chinadaily.com.cn](http://china.chinadaily.com.cn/a/201901/27/WS5c4d2c9ba31010568bdc6a44.html)*]

---
class: inverse, center, middle

# Guiding placement of health facilities using malaria criteria and interactive tool

---
# Access to healthcare and malaria transmission

---
# Access to healthcare and malaria transmission

---
# Distance / Travel time to health facilities

- Early diagnosis and treatment reduce death and transmission

- Many factors contribute to access to healthcare

- Distance or travel time to health facilities is important predictor to malaria prevalence *(e.g. Schoeps et al. 2011, Kizito et al. 2012)*

---
# Bunkpurugu-Yunyoo District, Ghana

.pull-left1[
- 1200 km `\(^2\)` , 150K populations

- 2 urban centers, 8 health facilities

- Multiple years of malaria surveys in 2010 - 2014

- Important predictors *(Amratia et al. 2019; Millar et al. 2018)*
  - Distance to health facilities (HF)
  - Distance to urban centers
]

.pull-right1[
<img src="../assets/img/byd-simplemap.png" width="1889" style="display: block; margin: auto;" />
]

---
# Expansion of health services?

- Community-based Health Planning and Services (CHPS)

- Run by trained community health officers (CHOs)
  
  - Assigned to community health post, services at doorstep
  
  - Treatment for malaria, acute respiratory diseases and other childhood illness
  
- Expand CHPS to all communities

---
# Prevalence vs Incidence

- Prevalence (0 to 1): Likelihood of a person in the area having malaria at a point in time

- Incidence: number of cases during a given period in the population of interest

- Available Data: Prevalence under 5 years old

- Decision criteria: Incidence (of all ages)

- Conversion from prevalence under five to incidence of all ages
  - Prevalence → Incidence not always linear, especially for older children and adults

---
# Objectives

- Determine the optimal locations for new health facilities based on district-wide malaria criteria:

1. Overall malaria prevalence of children under 5, 
  
  2. Overall malaria incidence of all ages, and 
  
  3. Average travel time to nearest health facilities
  
---
# Methods

- Three years of high transmission season data (2010 to 2012)

- ~ 5000 children under five

- 71 to 80 villages per year

- GAM with 5 predictors
  - Travel time to HF, distance to urban center, elevation, NDVI, log population density
  
- Use Genetic Algorithm to find optimal locations given number of health facilities and criteria

---
# Results

- See interactive visualizer and simulator created.
  - http://bit.ly/ben-hf-app
  
### Pitch...
<img src="../assets/img/valle-conbio-grab.PNG" width="1888" style="display: block; margin: auto;" />

---
class: inverse, center, middle

# Decision Support Tool to Predict Causes of Childhood Febrile Illness Using a Bayesian Model Approach

---
# Childhood Acute Febrile Illness (AFI)

.pull-left2[
- Malaria used to be the main cause of AFI in Africa

- With decline in malaria transmission, other etiologies such as bacterial infection can be increasingly prominent

- Difficulty in diagnosis, lack of test kit and prioritization of malaria can drive up misdiagnosis

- Predicting infection status based on symptoms, demographic variables and hematological variables
]

.pull-right2[
<img src="../assets/img/fever-clipart.png" width="351" style="display: block; margin: auto;" />
.source[*[kissclipart.com](https://www.kissclipart.com/cartoon-thermometer-clipart-thermometer-fever-clip-p9mkzm/)*]
]

---
# Statistical decision support tool

- Interactive tools that inform health practitioners using predicted probability (of infection)

- http://bit.ly/ben-afi-app

Ideal statistical decision support tool:
- Accommodate missing data
- User experience: Fast, small and flexible
- Accurate
- Informative and Interpretable

---
# Objectives

- Create model to simultaneously
  - Impute missing variables
  - Make predictions

- Modified Bayesian Model Averaging (BMA)

---
# Data

- 2016/2017
- 1500 children 1 to 15 years old

---
# Data

- Model diagnostic test outcomes using the recorded predictors

.pull-left[
Diagnostic Tests
- Malaria 
  - Rapid Diagnostic Test (RDT)
  - Microscopy
  - Taqman Array Card (TAC; Subsample)
- Blood and Urine Culture if deemed necessary
]

.pull-right[
Recorded Predictors
  - Age
  - Time of visit (season)
  - Body temperature
  - 30 clinical symptoms (Binary)
  - 10 Hematological parameters
  
😨 Missing values in 
  - 31/43 predictors (Most < 5%)
  - 17% individuals
]

---
# Model training

---
# Model prediction

---
# Performance vs other methods

---
# Key findings

- The modified BMA is as accurate as random forest without losing its interpretability
  - Better than that of stepwise logistic regressions

- Provides predictor importance for each diagnostic test

- Can be incorporated into a relatively responsive interactive decision support tool

---
# Key findings

- Poor predictive performance with Malaria TAC
  - Low sample size
  - Submicroscopic infection is asymptomatic → Symptoms are not useful predictors

- Poor predictive performance with Blood culture
  - Low sample size
  - Multiple types of pathogens – diverse signs and symptoms
  
---
# References

- Schoeps, Anja, et al. "The Effect of Distance to Health-Care Facilities on Childhood Mortality in Rural Burkina Faso", American Journal of Epidemiology 173, 5 (2011), pp. 492--498.
- Kizito, James, et al. "Improving access to health care for malaria in Africa: a review of literature on what attracts patients", Malaria Journal 11, 1 (2012), pp. 55.
- Weiss, D.J. et al. 2019. Mapping the global prevalence, incidence, and mortality of Plasmodium falciparum, 2000–17: a spatial and temporal modelling study. Lancet 394: 322–331.
- Mandal, S. et al. Mathematical models of malaria - a review. Malar J 10, 202 (2011).
- Millar, J. et al. 2018. Detecting local risk factors for residual malaria in northern Ghana using Bayesian model averaging. Malar. J. 17: 343.
- Amratia, P. et al. 2019. Characterizing local-scale heterogeneity of malaria risk: A case study in Bunkpurugu-Yunyoo district in northern Ghana. Malar. J. 18: 1–14.

---
# Acknowledgement

- Dr Denis Valle
- Committee members: Dr Nikolay Bliznyuk, Dr Rhoel Dinglasan, Dr Tom Hladish
- Collaborators: 
  - Dr Gordon Awandare, Dr Justin Stoler, Dr Nicholas Amoako and WACCBIP; 
  - Dr Benjamin Abuaku, Dr Paul Psychas, Noguchi Medical Research Institute and President's Malaria Initiative
- Labmates
- Family and friends

---
class: inverse, center, middle

# Thank you very much! 
Feedback, comments and questions? Email: kokbent [at] ufl.edu