Physician Versus Computer Coding of Verbal Autopsies

BACKGROUND

Most deaths in low and middle-income countries occur out of hospital and without medical attention and certification at the time of death. Hence, information on causes of death (COD) is lacking. In these settings, verbal autopsies (VAs), typically involving lay non-medical interviews of living family members or close associates of the deceased about the details of death, with subsequent assignment of COD by physician, can be used to estimate COD patterns.

Although VA is being commonly used for acquiring community-based COD data, its application and mode of assignment of COD may vary. The choice of physician-certified verbal autopsy (PCVA) coding versus computer-coded verbal autopsy methods (CCVA) has been widely debated. Both these methods have limitations. While PCVA methods suffer from inter and intra-observer differences in coding in addition to physician time consumption and expense, the accuracy of CCVA methods which are faster and less expensive than PCVA have not been assessed in different settings. A literature search yielded only one study that have systematically assessed the performance of four computer-coded verbal autopsy methods for COD assignment compared with physician coding of VAs on 24,000 deaths in low and middle-income countries (Miasnikof et al. 2015).

Here the investigators propose the first ever randomised control trial assessing PCVA versus CCVA at the population level.

STUDY DESIGN

This randomised control trial will be conducted in three states of India; a total of 12,500 deaths that occurred in the last five years will be collected via VA. In each district (five districts across the three states - see below), 50% of deaths are randomly selected for PCVA and the rest for CCVA. The COD distribution for both groups are then compared within each district. If the performance of PCVA and CCVA are comparable, the attained distributions should be similar.

Study districts in India (x number of anticipated VAs to be collected from approx. y number of households):

• Amravati, Maharashtra (2500 VAs from ~12,500 households)
• Anand and Kheda districts, Gujarat (5000 VAs from ~25,000 households)
• Sangrur and Mansa districts, Punjab (5000 VAs from ~25,000 households)

OBJECTIVES

• Primary Objective: To assess physician versus computer coding of VAs at the population level - Do computer algorithms perform as well as physician coding of VAs when determining the COD distribution at the population level.
• Secondary Objective: To assess the quality of physician versus lay surveyor VA data collection - Does the quality of symptom information collected or final COD assigned differ when a physician versus a lay person carries out the VA data collection.
• Tertiary Objective: To assess household preference of short VA questionnaire with narrative versus a long questionnaire without a narrative - Which VA instrument do households prefer.

DATA COLLECTION

Data collection for the study will proceed in two phases.

• Phase 1: Enumeration and VA (short questionnaire with narrative, and long questionnaire with narrative) by lay data collectors. VA questionnaires were designed based on World Health Organization (WHO) recommendations.
• Phase 2: Physician re-sampling for 50% of the VAs to be used for PCVA The data will be collected using a laptop/tablet in an electronic format in the local language.

Quality control/Quality Assurance:

1. Random 2-5% of cases will be re-interviewed by surveyors with the Quality Control Survey
2. An audio recording of the entire data collection will be done for quality control and monitoring.

ANALYSIS PLAN

• Primary Objective: COD for 50% of VAs will be assigned by trained physicians using the Million Death Study (MDS) physician coding system; this includes dual, independent coding of VA records, disagreements resolved by reconciliation, and remaining cases by adjudication by a third physician. The assignment of CODs will be in line with the international classification of disease version 10 (ICD-10). The remaining 50% of deaths will be independently assigned COD by five leading CCVA algorithms. If the performance of PCVA and CCVA are comparable, the attained COD distributions at the population level should be similar, this "similarity" will be assessed by calculating the Cause Specific Mortality Fraction (CSMF) Accuracy for each algorithm's COD assignment in each district against the CSMF of the PCVAs in the respective district.
• Secondary Objective: Half of the PCVAs in each district will be re-interviewed by physicians, and CODs will be assigned to these re-interviewed VAs using the MDS physician coding system described above. To assess concordance of COD assignment between lay person and physician collected VAs, the CSMF Accuracy will be calculated to determine concordance at the population level, while Sensitivity and Partial Chance Corrected Concordance (PCCC) will be calculated to assess similarity of COD assignment at the individual level.
• Tertiary Objective: Respondents cooperation with the short VA questionnaire with narrative versus the long questionnaire without a narrative will be assessed to determine which VA instrument households prefer. A focus group with surveyors to receive their feedback on using both instruments in the field will also be conducted.

The results of this study will be presented as three separate trials (one trial per state). The results of the first trial will be used for hypothesis generation.