Zambian HIV records expose data gaps that can distort AI

Zambia’s national HIV electronic health record system contains substantial gaps and inconsistencies that could allow artificial intelligence models to identify missing documentation as if it were clinical risk, according to a multi-facility assessment of 246,053 patient records from six public health facilities in Lusaka. Researchers evaluated the records across eight data-quality dimensions and used a machine-learning pipeline to examine which features mattered to models. Every facility fell below the study’s pragmatic 70% screening threshold on a composite index, with scores ranging from 53.5% to 58.3%. The findings show that several outcomes labeled as recorded in the database also depend on whether facilities tested patients or entered follow-up information, making apparent prevalence partly a measure of documentation practice.

The assessment used de-identified records from SmartCare, Zambia’s national HIV electronic record platform, covering Chawama First Level Hospital, Chelstone Urban Health Centre, George Urban Health Centre, Kanyama First Level Hospital, Matero Main Urban Health Centre and Matero Reference Hospital. Four facilities supplied 171,547 training records, while two containing 74,506 records were held out for evaluation. The cohort had a median age of 32 years, and 62.5% were female. Enrolments dated from 2003 to 2025, meaning many patients had little or no digitized clinical history. The study was retrospective and restricted to six Lusaka facilities, so its results may not represent rural, private or other district-level services.

Data completeness varied sharply by field and facility. Gender, age and tuberculosis treatment at enrolment were recorded for 99% to 100% of patients, while most physical measurements were also comparatively well captured. In contrast, education level was available for only 1.2% to 7.9% of patients across the six facilities, household income for 4.7% to 19.8%, and HIV enrolment stage for 21.1% to 24.6%. Past tuberculosis type, exit reason and the date of the last treatment interruption were effectively absent. Controlled values such as WHO stage and tuberculosis-treatment flags generally conformed to their expected vocabularies when present, but laboratory results were stored as free-text strings and required normalization. The pipeline also required 23 explicit schema workarounds because fields were absent or differently named between facilities, and date fields required tolerant parsing.

The most consequential problem was the collapse in coverage for features describing longitudinal care. Although the dataset included 246,053 patients, usable appointment and visit-history features could be calculated for only 30,295 patients, or 12.3%. Laboratory features for viral load, CD4 counts and tuberculosis tests were available for just 8,233 patients, or 3.3%. For most patients, therefore, the model received zero or training-median engagement values and imputed laboratory information rather than observed measurements. This left the analysis relying heavily on a one-time demographic snapshot, where flags for missing education or income could become influential substitutes for clinical information that was largely absent.

Some of the available longitudinal data also failed basic plausibility checks. Among the 30,295 patients with engagement information, 16,439, or 54.3%, had a maximum appointment-lateness value exceeding 730 days, even though the programme had existed since 2003; the largest value was 45,459 days, equivalent to about 124 years. Another 3,965 patients, or 13.1%, had negative lateness beyond 30 days, indicating recorded attendance more than a month before the scheduled appointment. These values entered the modelling pipeline unchecked because range limits were applied only to height and weight. Appointment lateness nevertheless ranked among the most important features in the shared pipeline, placing data-entry artifacts in a position where they could influence predictions.

The three 12-month outcomes were explicitly defined as recorded events rather than direct measures of incidence or population prevalence. Recorded tuberculosis treatment required a positive treatment field during the year after the index date, but could not distinguish incident disease from treatment begun earlier. Programme-recorded interruption required an explicit inactive status, and patients without such a record were treated as active; the resulting cohort estimate was 0.56%, compared with facility rates of 0.20% to 0.65%. Recorded unsuppressed viral load required a result of at least 1,000 copies per milliliter, while patients without a post-index test were coded as suppressed. Only 25,294 patients, or 10.3%, had at least one post-index viral-load test, and 14.1% of those tested had an unsuppressed result, whereas the full-cohort label produced a much lower prevalence of 1.45%.

Facilities with more testing consequently appeared to have more unsuppressed viral load, and their recorded tuberculosis-treatment rates ranged from 32.1% to 48.4%. These contrasts cannot establish that clinical risk genuinely differed between sites because coding, follow-up recording and testing intensity also varied. The study calculated exploratory Pearson correlations between facility-level missingness and outcome rates. Education missingness correlated with recorded viral suppression at r = −0.865, and HIV-stage missingness correlated with recorded tuberculosis treatment at r = +0.741. Because these estimates used only six facilities, however, they are ecological, hypothesis-generating associations with limited statistical power; they neither control for confounders nor show that any model prediction was caused by missing data.

Patient-level comparisons showed why the recorded subgroup was not representative. Of the full cohort, 12,766 patients had an education value, while 233,287 did not. Those with recorded education were about three times as likely to have pre-index engagement data, 34.7% compared with 11.1%, and about twenty times as likely to have a usable pre-index CD4 or viral-load result, 12.5% compared with 0.6%. They also had higher recorded rates of tuberculosis treatment, treatment interruption and unsuppressed viral load despite nearly identical median ages and sex distributions. In the companion model, education’s missing-data indicator had the highest mean SHAP value among 57 encoded features for recorded unsuppressed viral load. The present study’s separate permutation-importance analysis also ranked education first for interruption in treatment and unsuppressed viral load, corroborating the model’s sensitivity to incomplete records without proving the mechanism behind that sensitivity.

The researchers interpreted the patterns using Rubin’s missing-data taxonomy. Age and gender were nearly complete, making their residual missingness plausibly compatible with missing completely at random. Laboratory absence could often be explained by observable factors such as whether a test was due, making it plausibly missing at random. Education, income, enrolment stage, marital status and even valid body-mass index were missing in patterns structured by facility and enrolment period, a pattern consistent with missing not at random. The study used median imputation for numerical variables and most-frequent-value imputation for categorical variables, fitted only on the training facilities, followed by standardization and one-hot encoding. This single-imputation approach intentionally exposed the deployment-like weakness rather than repairing it, although multiple imputation with sensitivity analysis would better represent uncertainty for fields plausibly missing at random. The released code and assessment toolkit are intended to let teams holding equivalently structured SmartCare exports repeat the analysis.

The study’s broader conclusion is that trustworthy HIV AI must begin with data governance, not model selection alone. The authors proposed eight minimum requirements for resource-limited public-health systems, including standardized field definitions, explicit handling of unavailable values, automated outcome construction, and regular facility-level quality monitoring. The underlying patient-level data remain unavailable because of confidentiality and Ministry of Health governance policies, and patient consent was formally waived for the retrospective use of de-identified records. The work does not show that AI caused harm at any facility, nor does it establish an algorithm’s clinical benefit. It instead provides a reproducible measurement of a prerequisite: whether an HIV dataset contains enough complete, plausible, comparable and temporally stable information for a model to learn from clinical need rather than the administrative circumstances surrounding a patient’s record.

Subject of Research: Data quality and algorithmic fairness in machine learning for Zambia’s HIV electronic health records

Article Title: A multi-facility data quality assessment of electronic health records for trustworthy AI in HIV public health in Zambia

Article References: Phiri, J., Zimba, A., Njovu, C., Kashoka, A., Chiboma, I., Chiwele, M., Sinkala, T., Mutale, J., & Lumpa, M. (2026). A multi-facility data quality assessment of electronic health records for trustworthy AI in HIV public health in Zambia. Discover Artificial Intelligence, 6(1), Article 1208. https://doi.org/10.1007/s44163-026-02293-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02293-x

Keywords: artificial intelligence, HIV surveillance, electronic health records, data quality, algorithmic fairness, missing data, Zambia, machine learning, viral load, public health, multi-facility, data