va_healmod.knit

class: center, middle, inverse, title-slide

.title[
# <div class="line-block">Filling in the Gaps<br />
Verbal Autopsy and Cause of Death Assignment<br />
<br />
<br />
</div>
]
.author[
### Yue Chu </br></br> Jason Thomas </br></br> & Sam Clark
]
.institute[
### </br> HealMod Co-Laboratory Workshop </br>
]
.date[
### Oct. 1st, 2024
]

---

# Motivation

.center[<img src="img/tanz_coverage.png" style="width: 100%" />]

Source: Causes of Deaths From Community Settings in Tanzania.  Ministry of
Health (2024)

---
background-image: url("img/coverage_map.png")
background-size: contain

<div style="position: absolute; bottom: 0px">
<a href="https://www.who.int/data/data-collection-tools/score/dashboard#/">
Source: WHO Score Dashboard
</a>
</div>

---
# Capacity to Register CoD

.center[<img src="img/capacity.png" style="width: 75%" />]

Source: "Global report on health data systems and capacity, 2020"
[WHO](https://iris.who.int/bitstream/handle/10665/339125/9789240018709-eng.pdf?sequence=1)

---
# Implications & Solutions

* Knowledge of the leading causes of death at the community/population level
informs health polices and facilitates the evaluation of interventions.

--
  
  + It can take many years for civil registration and vital statistics systems
  to develop sufficient capacity.

* **Verbal Autopsy** -- interviewing the decedent's next of kin or caregiver
to gather data on the relevant symptoms and medical history.

+ Physicians can assign causes of death using the VA data.

+ Algorithms have also been developed to assign causes, providing
  an expedient solution for providing CoD information.

---
# openVA Team

.pull-left2[
<img src="img/openva.png" style="width: 75%" />
[openva.net](https::/openva.net)
]

.pull-right2[

* Multidisciplinary team led by Sam Clark (Dept. of Sociology, OSU)
  
* Support updates of the World Health Organization's Verbal Autopsy Instrument

* **Develop algorithms** for assigning causes to VA data

* Involved in VA data collection (with reference causes)

* Validation studies

* Software, training, and support VA data analysts
]

---
class: slide-font-27
# InSilicoVA

* InterVA -- popular VA algorithm that has been in service for 20+ years

* **InSilicoVA** [(McCormick, Li, Clark, et al., 2016)](https://www.tandfonline.com/doi/full/10.1080/01621459.2016.1152191)
-- developed to improve upon InterVA...
  
--

+ fully probabilistic model with a hierarchical structure that estimates
  
      - distribution of deaths by cause in the population (CSMF)
      
      - *probability of each cause* for individuals
      </br>(InterVA &rarr; "indeterminate")
      
--

+ provides uncertainty estimates at both the population and individual level

+ accounts for missing values in the data

---
# Validation Studies & Reports

[Natl. Cause of Death Validation Study (S. Africa)](https://www.tandfonline.com/doi/full/10.1080/16549716.2023.2285105)

* Can VA algorithms reproduce physician cause assignment?

* Individual Level: Agree on Top Cause (Top 3)

+ InSilicoVA -- 51.6% (73.8%)
  + InterVA5 -- 48.2% (70.9%)
  + Tariff2 -- 51.2% (*not available*)
  
--

* Population level: CSMF Accuracy

+ InSilicoVA -- 0.84
  + InterVA5 -- 0.81
  + Tariff2 -- 0.82

---
background-image: url("img/mrc_results.png")
background-size: contain

---
# Validation Studies & Reports

[Tanzanian MoH Report's Key Findings]()

* VA methodology works well; recommend scaling up to national level

* "The best performing CCVA algorithm was InSilicoVA with a CSMF
physician concordance of 83%.  InSilicoVA also has an advantage over
InterVA5 and Tariff2 in that it does not deliver undetermined results."
(p. 50)

---
# Moving Forward

There are various ways of improving VA cause assignment using computer coded
algorithms

* Improving algorithm performance

+ symptom dependence
  + domain adaptation models

* Use more information from the VA interview

* Growing the body of VA data with reference causes to train algorithms/models

+ build knowledge on epidemiological differences across space & time