Ziyaad Dangor

BACKGROUND Large language models (LLMs) are increasingly proposed for diagnostic support. However, validation of their performance in real-world multimodal inpatient care remains limited, particularly in low- and middle-income country (LMIC) hospital settings. OBJECTIVE This study aimed to evaluate the diagnostic accuracy and clinical reasoning of leading multimodal LLMs compared to routine clinical practice within a South African tertiary hospital. METHODS We conducted a retrospective evaluation of 10 leading multimodal LLMs using 539 complete inpatient cases from a South African tertiary public hospital. The dataset included radiology imaging, radiology reports, laboratory results, and clinical narratives. Expert clinician panels adjudicated 300 cases to establish reference diagnoses, differential diagnoses, and clinical reasoning. Model outputs and routine recorded ward diagnoses were scored using a panel-calibrated, three-model LLM Jury. RESULTS Mean LLM scores were tightly clustered despite 50-fold cost differences, and on average all models exceeded routine ward diagnostic performance (>99.9% bootstrap probability). In a senior clinician-adjudicated tie-breaker subset maximised for model–panel disagreement, a predefined Top-3 LLM ensemble exceeded the original expert-panel consensus with 97% bootstrap probability. Residual safety risks and model non-response motivate prospective evaluation with rigorous adjudication and clinical oversight. CONCLUSIONS Multimodal LLMs demonstrated high diagnostic performance in a complex, real-world LMIC inpatient setting, surpassing recorded routine ward diagnoses in 60-80% of cases depending on the model. Our finding that even the cheapest models outperform the routine ward diagnoses and are comparable to the best models suggests that affordable AI clinical decision support in resource-constrained environments is now worth serious consideration.

Background: Though preventable, burns are the leading cause of trauma‐related death in children under 4 years in South Africa and account for almost half of all pediatric trauma cases at Chris Hani Baragwanath Academic Hospital. We investigated the impact of an education‐based pediatric hot water burns prevention program on the knowledge and behavior of caregivers in Soweto. Methods: We conducted a quasi‐experimental nonequivalent controlled before‐and‐after intervention study with a 7‐month follow‐up. Households were randomly selected within purposively identified high‐burden burns clusters. At baseline, household assessments and caregiver surveys captured burn history, burn risk, first aid knowledge, and household injury risk. Subsequently, all caregivers received structured education on pediatric burns, prevention measures, and first aid. At follow‐up, the same households and individuals were reassessed to evaluate for knowledge and behavior change. Results: A total of 242 intervention and 52 control households were assessed. There were no significant differences between the control and intervention households. Most households were severely overcrowded (70.8%, 208/294). Only 7.1% (21/294) had geysers, with 89.8% (264/294) dependent on manually boiled water for bathing, despite most households being electrified (97.9%, 288/294). Relative to baseline, there was a 30.8% (95% CI 22.2%–40.9%) increase in knowledge on the correct first aid for hot water burns postintervention. A 43.3% (95% CI 36.5%–50.5%) increase in stove tops being kept > 1.4 m off the ground, 40.3% (95% CI 30.0%–51.4%) increase in appropriately checking the temperature of bath water, 38.3% (95% CI 29.1%–48.4%) increase in ensuring pot handles are out of reach, and 23.5% (95% CI 15.8%–33.6%) decrease in handling of hot beverages while holding a child were also noted postintervention (all exact McNemar p < 0.001). There was no change in the number of households with childfree safe zones postintervention. Conclusions: Caregiver education was associated with significant improvements in caregiver knowledge and selected safety practices. Structural constraints, such as overcrowding, may limit the adoption of some positive behavioral changes, underscoring the need for additional system level intervention.

A bs t r act: Accurately attributing causes of death is vital for global health, yet fewer than 5% of deaths in resource-constrained regions are medically certified. To assign causes to these unlabeled deaths at scale, practitioners traditionally rely on verbal autopsy, using supervised statistical models to classify based on structured survey data. However, modern mortality surveillance increasingly collects rich, unstructured multimodal data, such as free-text caregiver narratives and postmortem diagnostics, which traditional supervised statistical models struggle to seamlessly integrate. In this paper, we present a comprehensive, multimodal benchmark for cause-of-death classification using data from the Child Health and Mortality Prevention Surveillance (CHAMPS) network, a unique surveillance platform spanning nine countries across South Asia and Sub-Saharan Africa. Using this dataset, we introduce an evaluation framework designed to rigorously assess diagnostic reasoning, moving beyond traditional metrics that fail to capture complex clinical realities. We demonstrate the utility of this benchmark by evaluating zero-shot large language models against supervised baselines across various data modalities. Our results reveal distinct differences in how these modeling approaches synthesize unstructured medical evidence. This benchmark provide a rigorously defined resource for assessing clinical reasoning in next-generation mortality surveillance.

Background: Klebsiella pneumoniae is a leading cause of sepsis in young infants. We evaluated the association of invasive K. pneumoniae disease (iKPnD) in relation to antigen-specific immunoglobulin G (IgG) and serum bactericidal activity (SBA) to four polysaccharide capsular (K) serotypes and five lipopolysaccharide (O) serotypes, as well as IgG to MrkA, in infants less than 90 days of age. Methods: We conducted a retrospective case-control study in Kenyan and South African infants with blood culture-confirmed iKPnD. Serotype-specific antigen IgG concentrations of cases were compared with hospitalised controls without iKPnD. Geometric mean concentrations (GMCs) were estimated, and scaled covariate-adjusted models were used to estimate risk reduction over a grid of antibody concentrations. Results: Transplacental transfer of IgG against various K. pneumoniae antigens increased with advancing gestational age. Serum IgG GMCs (expressed in RLU/mL) to disease-causing homotypic K- or O-serotypes were lower in cases compared with controls for anti-K2 (396 [95%CI: 250-628] vs 660 [95%CI: 562-776]), anti-K25 (396 [95%CI: 251-623] vs 1170 [95%CI: 988-1385]), anti-K149 (327 [95%CI: 204-521] vs 492 [95%CI: 435-557]); as well as anti-O1αβ,2α IgG (1282 [95%CI: 782-2101] vs 2250 [95%CI:1904-2658]). Furthermore, overall anti-MrkA IgG was lower in cases (945; 95%CI: 757-1179) compared with controls (1610; 95%CI: 1378-1880). SBA titres (expressed as IC50) did not differ between case and controls by K types, but were lower for O1αβ,2β in cases (27; 95%CI: 12-63 vs. 136; 95% CI: 84-221). Conclusion: Our findings provide preliminary evidence that low antibodies against three of four K-antigens, O1αβ,2β and MrkA are inversely associated with iKPnD, and should be explored as potential vaccine antigens.

Background: Post-market vaccine safety surveillance and research are essential to detect and evaluate rare adverse events following immunization (AEFIs) not identified in pre-licensure clinical trials, such as myocarditis post-COVID-19 vaccination. AEFI surveillance infrastructure varies between jurisdictions. Factors that enable or hinder AEFI investigation across diverse settings are not well understood. Objective: To examine clinical, social, and structural/system factors that enable AEFI case identification, reporting, data and biosample collection in high-income countries (HICs) and low-and middle-income countries (LMICs). Methods: We conducted a qualitative study informed by Interpretive Description in Kenya, South Africa, and Canada. Participants were recruited using purposive and snowball sampling from three groups: (1) key informants with leadership roles in AEFI surveillance and research, (2) healthcare workers, research and laboratory staff involved in AEFI investigation and research, and (3) past participants in vaccine studies. Data were collected through semi-structured interviews and focus group discussions. Thematic analysis was conducted. Results: Eighteen key informants, 47 healthcare workers and research staff, and 27 past research participants were enrolled. Facilitators and barriers were identified across two domains: AEFI surveillance and investigation, and vaccine safety research participation. AEFI surveillance and investigation were shaped by trust in system responsiveness, beliefs and awareness around AEFI reporting, digital innovation, and implementation gaps. Research participation was shaped by altruism and social influence, logistics and research infrastructure, and institutional policies and privacy. Facilitators to AEFI surveillance and research included established policies and procedures for AEFI surveillance, digital tools, emphatic communication, and convenient research processes. Barriers included limitations in workforce capacity, diagnostic testing and funding (especially in LMICs), and cumbersome research approval processes. Conclusions: Findings underscore the need for both globally harmonized standards and locally tailored strategies to support vaccine safety surveillance and research in LMICs and HICs that are patient/participant-centered and include sustainable investments in infrastructure and workforce capacity.

Coronavirus disease 2019 (COVID-19) following Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) infection caused a 41-month global public health emergency (January 2020–May 2023). Although the pandemic has subsided, there is ongoing concern about emerging viral variants [1]. Furthermore, knowledge gaps persist regarding paediatric COVID-19 clinical presentation, treatment and outcomes in high HIV prevalence settings, with […]

Background: There is a paucity of data on gastrointestinal colonisation with Klebsiella pneumoniae (KPn) in African adults, although it is a known risk factor for developing KPn invasive disease (KPn-ID). Objectives: We investigated the risk factors for KPn gastrointestinal colonisation among a cohort of hospitalised adults without KPn-ID in South Africa, and described their clinical outcomes. Method: A cohort of hospitalised adults without KPn-ID was enrolled between 15 May 2023 to 14 May 2024 across three hospitals in Johannesburg, South Africa. Study participants had rectal swab cultures performed once, at enrolment, to determine the presence or absence of KPn gastrointestinal colonisation, and were followed up until in-hospital death or discharge. Results: Among the cohort of hospitalised adults without KPn-ID (n = 651), the rate of KPn gastrointestinal colonisation was 34.4% (224/651). Risk factors for KPn colonisation were peptic ulcer disease (adjusted odds ratio [aOR] 7.75; 95% confidence interval [CI]: 1.34-44.78), a central venous catheter (aOR 3.07; 95% CI: 1.00-9.40), a presumed healthcare-associated infection (pHAI) (aOR 3.13; 95% CI: 1.06-9.28), and length of stay more than seven days prior to enrolment (aOR 1.84; 95% CI: 1.23-2.75). There was no difference in in-hospital case fatality risk (CFR) between participants with (3.1%; 7/224) and without (3.0%; 13/427, p = 0.955) KPn colonisation. Conclusion: KPn gastrointestinal colonisation rates among hospitalised adults without KPn-ID were higher than that of high-income settings and also varied between study sites. Length of hospital admission prior to swabbing and pHAI were associated with KPn colonisation, suggesting that infection prevention control measures play a significant role in KPn colonisation. Contribution: Large prospective cohort studies and community surveillance studies are required in settings such as ours to further investigate KPn colonisation dynamics and explore preventive strategies against KPn-ID.