Sara Khalid

Rare diseases (RD) are individually uncommon but collectively affect millions of people. Yet national estimates of their prevalence, especially across ethnic groups, are limited, and how the COVID-19 pandemic affected people with RD across ethnic groups is unknown. We aimed to estimate the prevalence of 406 RD by ethnicity, and to describe recorded COVID-19 infection across these RD and ethnic groups, in England. We conducted a cross-sectional study within the National Health Service (NHS) England Secure Data Environment, accessed via the British Heart Foundation (BHF) Data Science Centre’s CVD-COVID-UK/COVID-IMPACT Consortium, linking primary care, hospital, and mortality records for individuals alive on 31 July 2023. We calculated age- and sex-standardised prevalence for 406 RD, and recorded COVID-19 infection rates for the overall group, for each RD, and across the six-group Office for National Statistics (ONS) and 19-group NHS ethnicity classifications. Of 62.5 million people, 1.05 million (1.7%) had at least one of the 406 RD. Recorded COVID-19 infection was higher in the RD group than in the non-RD population (32.8% vs. 29.0%). Within the RD group, infection rates varied by ethnicity (e.g. 27.7% in the Pakistani and 23.8% in the Black African groups, versus 33.7% in the White British group) and by specific RD (24.6–41.2% among the 25 most common). Age- and sex-standardised prevalence also varied widely across ethnic groups, with substantial heterogeneity emerging only when broad categories were disaggregated into the more granular NHS classifications (for example, markedly elevated discoid lupus erythematosus prevalence in Black Caribbean but not Black African individuals, obscured within the aggregate “Black” category). This national, ethnicity-resolved map shows that rare-disease prevalence varies substantially across ethnic groups, with differences largely hidden within broad categories, and that recorded COVID-19 infection among people with RD varied across ethnic groups, with the highest rates observed in the ONS White and NHS White British groups. These findings provide a baseline to support earlier, fairer diagnosis and more targeted care.

Background: Rising temperatures are recognized as an environmental determinant of child health, but evidence linking heat exposure to child undernutrition has used heterogeneous temperature metrics and exposure windows. We investigated the association between short-term exposure to high temperatures and weight-for-height z-score (WHZ) among children under 5 years in low- and middle-income countries (LMIC). Methods: We pooled Demographic and Health Surveys conducted since 2015 from 36 LMICs. Short-term heat exposure was defined as the average daily maximum temperature (ADMT) during the 61 days preceding the median interview date. We estimated associations between ADMT and WHZ using linear regressions with country-fixed effects, adjusting for covariates identified through a causal framework, and evaluated effect modification by child age and household wealth. Uniform hypothetical temperature increases (+1°C to +5°C) were used to estimate excess wasting. Findings: Each 1°C increase in ADMT was associated with a 0·034 SD reduction in WHZ (95% CI -0·037 to -0·030). The association was strongest among infants, with reductions of up to 0·045 SD per 1°C (95% CI -0·051 to -0·038), attenuating with age, and did not differ by wealth. A hypothetical 2°C increase in ADMT corresponded to an estimated 1·6 million additional wasted children, increasing to 4·3 million under a 5°C increase. Interpretation: Recent heat exposure was consistently associated with poorer acute nutritional status among young children in LMICs, particularly during the first two years of life. Continued global warming could substantially increase childhood wasting, underscoring the need to incorporate climate adaptation into child nutrition policy and programming.

Background: Climate-exacerbated flooding triggers global public health crises, costing lives, livelihoods, and jeopardising fragile healthcare systems. Despite well-established disaster management frameworks, the timely communication of rapid, locally actionable information remains lacking—leaving vulnerable populations at risk. We developed and operationally validated an open-access monitoring system across six South Asian nations, reporting the public health impact of the 2025 floods and executing live daily nowcasting during the July 2026 monsoon emergency to better understand who is in harms’ way, democratise anticipatory disaster intelligence, and support healthcare resilience. Methods: We engineered a multi-stakeholder co-designed ensemble pipeline for flood-health intelligence by fusing observations from six synthetic aperture radar and optical satellites with hydrological, demographic, agricultural, transportation, and infrastructure datasets. We estimated cumulative exposure for children and adults, alongside functional disruption to healthcare facilities, built-up areas, road networks, and cropland across remote, rural, and urban areas of Pakistan, India, Bangladesh, Bhutan, Nepal, and Sri Lanka. The 2025 analysis was benchmarked against publicly-available official surveillance reports and independent field-validation. The 2026 deployment is ongoing live, externally field-validated against 8 ground-sensor stations along the Indus River system. Findings: Our analysis revealed a critical gap in current surveillance. In 2025, continuous monitoring identified 257.67 million exposed individuals (29.1% in rural areas) including 89.3 million children under 18 (34.6%). A five-fold increase (72.19 million additional people) was found over the 18.97 million captured in official episode-level reports for Pakistan, Bangladesh, and Sri Lanka. Over 460, 000 km² of agricultural land was exposed (18.5 % to >2m deep water), having significant implications for the region’s predominantly agrarian economy.Across publicly unmonitored regions and flood episodes, an estimated 237 million flood-exposed people were left in a data vacuum, completely missing from publicly-available mapping. Functionally, 22,818 low-lying healthcare facilities were inundated (3,041 functionally disrupted), severing last-mile care for 46.79 million people, including 28.54 million exposed to deep-water flooding (>2m). In each country, access to at least 5% healthcare facilities was disrupted. Our rapid monitoring of the July–August 2026 crisis has already reported 11,905 km² of inundation, 92,140 km of submerged roads, 5.19 million exposed people including 1.6 million children (30.8%), and disruption to 1,900 basic health units, community clinics, pharmacies, and hospitals. Interpretation: This first region-wide, multi-year study demonstrates that current global disaster response relies on severely incomplete or inaccessible data, effectively disenfranchising the world's most underserved and climate-vulnerable populations from healthcare when most needed. Rapid, reliable, and scalable precision intelligence is not a technical luxury but a necessary and achievable instrument for health equity and climate justice. To counter 21st-century global health threats, we recommend: (1) prioritising anticipatory planning for children and vulnerable communities via national action plans and community toolkits, (2) democratising disaster data through digital public infrastructure to inform e

Objectives: We aimed to assess the risk of incident autoimmune and inflammatory conditions during the post-acute period of COVID-19. Design: Descriptive network cohort study. Setting: Electronic health records from the UK and Dutch primary care, Norwegian linked health registry, hospital records of specialist centres in Spain, France and Korea and healthcare claims from Estonia and the USA. Participants: We followed individuals between September 2020 and the latest available data from day 91 after a SARS-CoV-2 negative test (comparator) or a COVID-19 record (exposed patients, ie assessing patients during the post-acute phase). We further established a reinfection cohort (any further COVID-19 record among the exposed patients). We followed patients until an outcome, end of study period, death, day 365 or an infection (comparator only) or reinfection (exposed patients only). Main outcome measures: We assessed postural orthostatic tachycardia syndrome (POTS) diagnoses/symptoms, myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS) diagnoses/symptoms, multi-inflammatory syndrome (MIS) and several autoimmune diseases (rheumatoid arthritis (RA), juvenile idiopathic arthritis (JIA), systemic lupus erythematosus (SLE), inflammatory bowel disease (IBD) and type 1 diabetes mellitus (T1DM)). Meta-analysed crude incidence rate ratios (IRRs) of outcomes after COVID-19 versus negative testing and after reinfection versus a previous COVID-19 record yield the ratios of respective absolute risks of each assessed outcome. We performed subgroup analyses by age, sex and predominant variant periods. Results: We included 2 521 812 individuals with a first COVID-19 record, 4 233 145 with a first negative test and 135 551 with a reinfection. Age and sex were largely comparable between exposure groups with a shorter follow-up for the reinfection cohorts. After COVID-19 compared with test-negative patients and equally after reinfection compared with previous COVID-19 patients, we did not observe increased rates for all outcomes and all subgroup analyses. Counts of MIS and JIA were too small for meta-analyses. Conclusions: In our descriptive meta-analyses of crude IRRs among databases from various countries and settings, we did not observe increased rates of incident POTS, ME/CFS, RA, IBD, SLE and T1DM in COVID-19 versus test-negative or reinfection versus COVID-19 during the first 9 months of the post-acute phase of COVID-19 or reinfection (>90 days postinfection until month 12). Since causal interpretation cannot be made from this study, further causal research is warranted.

The capacity to monitor shark biodiversity is vital for conservation. Baited Remote Underwater Video Systems (BRUVS) are an increasingly popular and cost-effective method to characterize their biodiversity, abundance, and distribution. However, BRUV monitoring remains hindered by extensive manual video annotation and a lack of accessible tools to generate datasets for ecological analysis.To overcome these barriers, we developed a suite of automatic tools for shark detection and taxonomic classification. We integrated an object detection model (SharkTrack) with hierarchical species classification (Shark Detector) trained on 264,712 images of 309 shark species, in two open platforms—an R package (sharkDetectoR) and a lightweight desktop application (SharkByte)—making the updated model package an accessible tool across computational environments and user expertise.Tested on 46,332 holdout images and 13.9 hours of BRUV footage from the Main Hawaiian Islands and Palau, these tools achieved 92% species-level accuracy as well as 94.7% detection accuracy, and reduced annotation time by up to 90%.These findings underscore the value of combining scalable artificial intelligence to advance shark biodiversity monitoring. By substantially reducing manual annotation demands while maintaining high detection and classification performance, this framework offers a practical solution to speed up BRUV post-processing and filtering of digital observations. At the same time, it demonstrates automated methods as crucial tools for conservation and management, especially to address species-specific data deficiencies in shark research.

Indoor air quality (IAQ) has emerged as a critical public health concern in rapidly urbanizing cities like Lahore, Pakistan, where ambient pollution levels regularly exceed WHO safety thresholds during annual smog seasons. This study aims to quantify the independent and combined effects of indoor plants and natural ventilation on IAQ through a quasi-experimental design, with implications for affordable IAQ management in smog-affected urban contexts. Eight office spaces across Beaconhouse National University were monitored over three months (July-September) using Awair Element consumer-grade sensors measuring CO 2 , PM 2.5 , PM 10 , and TVOCs at five-minute intervals. Six air-purifying plant species-Dracaena trifasciata, Aloe vera, Epipremnum aureum, Chlorophytum comosum, Rhapis excelsa, and Chamaedorea seifrizii-were deployed to assess remediation capacity under varying ventilation scenarios. Results showed that natural ventilation was highly effective, with windowed spaces exhibiting 96.2% lower PM 2.5 (2 µg/m 3 vs. 53 µg/m 3) and 70.5% lower TVOCs (535 ppb vs. 1,814 ppb) compared to sealed rooms. Conversely, plants in non-ventilated spaces achieved substantial PM 2.5 reductions (96.7%) but paradoxically increased TVOCs by 63.3%, likely due to human activity overwhelming plant metabolic capacity. Temporal analysis revealed that CAM plants effectively reduced nighttime pollutants when occupancy ceased, though daytime human activity in sealed offices generated pollutants faster than plants could metabolize them. These findings demonstrate that phytoremediation effectiveness critically depends on adequate ventilation and appropriate space-to-plant ratios. The study concludes that meaningful improvements in IAQ in Pakistani urban contexts require integrated strategies combining ventilation, strategic plant deployment, and effective policies on human presence in enclosed spaces, particularly during Lahore's severe smog season.

Background: As the global focal point of the intensifying type 2 diabetes mellitus (T2DM) epidemic, China bears a formidable public health burden. While wildfire PM2.5 is a pivotal climate-related health threat, its long-term impact on T2DM incidence remains poorly characterized. Notably, the potential for risk heterogeneity across the life course remains largely unexplored.Methods: Utilizing a population-based cohort of over 9 million individuals in China (2017-2024), this study investigated the life-course effects of wildfire PM2.5 exposure across the full T2DM disease spectrum, ranging from longitudinal changes in fasting blood glucose (FBG) and pre-T2DM to incident T2DM and its complications. Exposure-response relationships were modeled, and the T2DM burden attributable to wildfire PM2.5 exposure was estimated within a counterfactual framework.Findings: Overall, each 1 μg/m3 increment in wildfire PM2.5 and ambient PM2.5 (as a benchmarking) was associated with significantly elevated risk of incident T2DM (HR and 95% CI: 1.371 [1.321, 1.423] vs 1.004 [1.003, 1.004]) and its complications (1.971 [1.904, 2.040] vs 1.008 [1.007, 1.008]). Similarly, wildfire PM2.5 exposure was linked to an elevated risk of pre-T2DM (OR and 95% CI: 1.446 [1.430, 1.462] vs 1.004 [1.004, 1.004]) and an increased annual change in FBG (β and 95% CI: 0.034 [0.033, 0.036] vs 0.000 [0.000, 0.000]). This significantly pronounced association of wildfire PM2.5 was consistently observed across the young, middle and older adults, and presented monotonic upward trends across the full T2DM spectrum. Under a counterfactual scenario (5th percentile of exposure), the population attributable fractions for T2DM, T2DM complications, and pre-T2DM were 14.05% (12.56%, 15.53%), 18.79% (17.23%, 20.36%), and 7.27% (6.91%, 7.63%), respectively.Interpretation: Long-term exposure to wildfire PM2.5 is associated with increased risk across the full T2D spectrum and life course. Its metabolic threat substantially exceeds ambient PM2.5, and this elevated risk persists even at lower exposure levels.

Objective: Large Language Models (LLMs) hold the potential for clinical task-shifting by processing unstructured clinical text, enabling tasks such as clinical concept extraction and medical question answering from electronic health records. If implemented reliably, such approaches may benefit over-burdened healthcare systems, particularly in resource-limited settings and for traditionally overlooked populations, provided that local fine-tuning is supported by appropriate clinical and technical expertise. However, this powerful technology remains largely understudied in real-world contexts, particularly in the Global South. This study aims to assess whether openly available LLMs can be used reliably for processing medical notes in real-world settings in South Asia. Methods: We used publicly available LLMs to parse de-identified clinical notes from a large electronic health records (EHR) database in Pakistan, containing hospital records for 8.2 million patients. ChatGPT (GPT-3.5) as a general-purpose LLM, and GatorTron (base), BioMegatron, BioBert and ClinicalBERT as medical LLMs were evaluated when applied to these data, after fine-tuning them with (a) publicly available clinical datasets namely Informatics for Integrating Biology & the Bedside (I2B2) and National NLP Clinical Challenges (N2C2) for medical concept extraction (MCE) and emrQA for medical question answering (MQA), and (b) the local Pakistani de-identified EHR dataset, which includes inpatient Discharge Summaries (DS) and Subjective, Objective, Assessment, and Plan (SOAP) notes, as detailed in this paper. MCE models were applied to these clinical notes using both 3-label and 9-label formats, while MQA models were applied to medical questions. Internal and external validation performance was measured for (a) and (b) using F1 score, precision, recall, and accuracy for MCE and BLEU and ROUGE-L, which measure lexical and sequence similarity, for MQA. Results: When clinical LLMs were not fine-tuned on the local EHR dataset, their performance during external validation on local data was notably poorer compared to internal validation on the dataset used for fine-tuning, with reductions of at least 15% in F1 scores for MCE and 35% in ROUGE-L and BLEU scores for MQA tasks. This suggests potential bias and highlights the inability of the medical LLMs to reliably handle the data distribution of the local population without further fine-tuning and adaptation. This trend persisted across two distinct natural language processing tasks: concept extraction and question answering, spanning a spectrum of task complexities. However, fine-tuning the LLMs with local EHR data significantly improved model performance across both tasks, yielding a 7.5% to 15% increase in the F1 score for MCE and a 27% to 53% increase in ROUGE-L and BLEU scores for MQA. Notably, ChatGPT, as a general-purpose LLM, stood out as an exception, demonstrating superior performance across all measured metrics on the local dataset compared to the publicly available dataset, with improvements ranging from 3% to 17% on the local EHR dataset, even without fine-tuning on the local data. Conclusions: Publicly available LLMs, predominantly trained on data from high-income regions, were found to be unreliable when applied in a real-world clinical setting in Pakistan. Fine-tuning them with local EHR data and regional clinical contexts improved their reliability, demonstrating a feasible adaptation strategy that is substantially less resource-intensive than training large language models from scratch. Close collaboration between local clinical and technical experts to curate and leverage more representative, inclusive, and unbiased medical datasets, can play a crucial role in further ensuring reliability of LLMs for resource-limited, overburdened settings, to be used in ways that are safe, fair, and beneficial for all.

Drawing on the experiences and lessons learned from researchers based in low- and middle-income countries (LMICs) that leverage generative artificial intelligence (GenAI) technologies to address socio-economic challenges, we showcase the considerable potential to use GenAI to accelerate the progress towards achieving some of the Sustainable Development Goals, as well as considerable obstacles for creating locally adapted AI tools for fair development in LMICs. An expanded evidence base on GenAI in resource-limited settings is crucial for policymakers to understand opportunities and risks, while rights-based safeguards against AI harms can be strengthened by the lived experiences of local projects.