AI Estimation of Gestational Age from Blind Ultrasound Sweeps in Low-Resource Settings

Statistical Characterisation of Fetal Anatomy in Simple Obstetric Ultrasound Video Sweeps

OBJECTIVE

We present a statistical characterisation of fetal anatomies in obstetric ultrasound video sweeps where the transducer follows a fixed trajectory on the maternal abdomen.

METHODS

Large-scale, frame-level manual annotations of fetal anatomies (head, spine, abdomen, pelvis, femur) were used to compute common frame-level anatomy detection patterns expected for breech, cephalic, and transverse fetal presentations, with respect to video sweep paths. The patterns, termed statistical heatmaps, quantify the expected anatomies seen in a simple obstetric ultrasound video sweep protocol. In this study, a total of 760 unique manual annotations from 365 unique pregnancies were used.

RESULTS

We provide a qualitative interpretation of the heatmaps assessing the transducer sweep paths with respect to different fetal presentations and suggest ways in which the heatmaps can be applied in computational research (e.g., as a machine learning prior).

CONCLUSION

The heatmap parameters are freely available to other researchers (https://github.com/agleed/calopus_statistical_heatmaps).

Overview of the Global and US Burden of Preterm Birth

Preterm birth (PTB) is the leading cause of morbidity and mortality in children globally, yet its prevalence has been difficult to accurately estimate due to unreliable methods of gestational age dating, heterogeneity in counting, and insufficient data. The estimated global PTB rate in 2020 was 9.9% (95% confidence interval: 9.1, 11.2), which reflects no significant change from 2010, and 81% of prematurity-related deaths occurred in Africa and Asia. PTB prevalence in the United States in 2021 was 10.5%, yet with concerning racial disparities. Few effective solutions for prematurity prevention have been identified, highlighting the importance of further research.

Deep learning to estimate gestational age from fly-to cineloop videos: A novel approach to ultrasound quality control

OBJECTIVE

Low-cost devices have made obstetric sonography possible in settings where it was previously unfeasible, but ensuring quality and consistency at scale remains a challenge. In the present study, we sought to create a tool to reduce substandard fetal biometry measurement while minimizing care disruption.

METHODS

We developed a deep learning artificial intelligence (AI) model to estimate gestational age (GA) in the second and third trimester from fly-to cineloops-brief videos acquired during routine ultrasound biometry-and evaluated its performance in comparison to expert sonographer measurement. We then introduced random error into fetal biometry measurements and analyzed the ability of the AI model to flag grossly inaccurate measurements such as those that might be obtained by a novice.

RESULTS

The mean absolute error (MAE) of our model (±standard error) was 3.87 ± 0.07 days, compared to 4.80 ± 0.10 days for expert biometry (difference -0.92 days; 95% CI: -1.10 to -0.76). Based on simulated novice biometry with average absolute error of 7.5%, our model reliably detected cases where novice biometry differed from expert biometry by 10 days or more, with an area under the receiver operating characteristics curve of 0.93 (95% CI: 0.92, 0.95), sensitivity of 81.0% (95% CI: 77.9, 83.8), and specificity of 89.9% (95% CI: 88.1, 91.5). These results held across a range of sensitivity analyses, including where the model was provided suboptimal truncated fly-to cineloops.

CONCLUSIONS

Our AI model estimated GA more accurately than expert biometry. Because fly-to cineloop videos can be obtained without any change to sonographer workflow, the model represents a no-cost guardrail that could be incorporated into both low-cost and commercial ultrasound devices to prevent reporting of most gross GA estimation errors.

Obstetric and neonatal outcomes in South Africa

BACKGROUND

Background epidemiologic population data from low- and middle-income countries (LMIC), on maternal, foetal and neonatal adverse outcomes are limited. We aimed to estimate the incidence of maternal, foetal and neonatal adverse outcomes at South African maternal vaccine trial sites as reported directly in the clinical notes as well as using the 'Global Alignment of Immunization Safety Assessment in Pregnancy' case definitions (GAIA-CDs). GAIA-CDs were utilized as a tool to standardise data collection and outcome assessment, and the applicability and utility of the GAIA-CDs was evaluated in a LMIC observational study.

METHODS

We conducted a retrospective record review of maternity and neonatal case records for births that occurred in Soweto, Inner City- Johannesburg and Metro-East Cape Town, South Africa, between 1st July 2017 and 30th June 2018. Study staff abstracted data from randomly selected medical charts onto standardized study-specific forms. Incidence (per 100,000 population) was calculated for adverse maternal, foetal and neonatal outcomes, which were identified as priority outcomes in vaccine safety studies by the Brighton Collaboration and World Health Organization. Outcomes reported directly in the clinical notes and outcomes which fulfilled GAIA-CDs were compared. Incidence of outcomes was calculated by combining cases which were either reported in clinical notes by attending physicians and/ or fulfilled GAIA-CDs.

FINDINGS

Of 9371 pregnant women enrolled, 27·6% were HIV-infected, 19·9% attended antenatal clinic in the 1st trimester of pregnancy and 55·3% had ≥1 ultrasound examination. Fourteen percent of women had hypertensive disease of pregnancy, 1·3% had gestational diabetes mellitus and 16% experienced preterm labour. There were 150 stillbirths (1·6%), 26·8% of infants were preterm and five percent had microcephaly. Data available in clinical notes for some adverse outcomes, including maternal- & neonatal death, severe pre-eclampsia/ eclampsia, were able to fulfil GAIA-CDs criteria for all of the clinically-reported cases, however, missing data required to fulfil other GAIA-CD criteria (including stillbirth, gestational diabetes mellitus and gestational hypertension) led to poor correlation between clinically-reported adverse outcomes and outcomes fulfilling GAIA-CDs. Challenges were also encountered in accurately ascertaining gestational age.

INTERPRETATION

This study contributes to the expanding body of data on background rates of adverse maternal and foetal/ neonatal outcomes in LMICs. Utilization of GAIA-CDs assists with alignment of data, however, some GAIA-CDs require amendment to improve the applicability in LMICs.

FUNDING

This study was funded by Pfizer (Inc).

Artificial intelligence, ChatGPT, and other large language models for social determinants of health: Current state and future directions

This perspective highlights the importance of addressing social determinants of health (SDOH) in patient health outcomes and health inequity, a global problem exacerbated by the COVID-19 pandemic. We provide a broad discussion on current developments in digital health and artificial intelligence (AI), including large language models (LLMs), as transformative tools in addressing SDOH factors, offering new capabilities for disease surveillance and patient care. Simultaneously, we bring attention to challenges, such as data standardization, infrastructure limitations, digital literacy, and algorithmic bias, that could hinder equitable access to AI benefits. For LLMs, we highlight potential unique challenges and risks including environmental impact, unfair labor practices, inadvertent disinformation or "hallucinations," proliferation of bias, and infringement of copyrights. We propose the need for a multitiered approach to digital inclusion as an SDOH and the development of ethical and responsible AI practice frameworks globally and provide suggestions on bridging the gap from development to implementation of equitable AI technologies.

Fetal biometry and amniotic fluid volume assessment end-to-end automation using Deep Learning

Fetal biometry and amniotic fluid volume assessments are two essential yet repetitive tasks in fetal ultrasound screening scans, aiding in the detection of potentially life-threatening conditions. However, these assessment methods can occasionally yield unreliable results. Advances in deep learning have opened up new avenues for automated measurements in fetal ultrasound, demonstrating human-level performance in various fetal ultrasound tasks. Nevertheless, the majority of these studies are retrospective in silico studies, with a limited number including African patients in their datasets. In this study we developed and prospectively assessed the performance of deep learning models for end-to-end automation of fetal biometry and amniotic fluid volume measurements. These models were trained using a newly constructed database of 172,293 de-identified Moroccan fetal ultrasound images, supplemented with publicly available datasets. the models were then tested on prospectively acquired video clips from 172 pregnant people forming a consecutive series gathered at four healthcare centers in Morocco. Our results demonstrate that the 95% limits of agreement between the models and practitioners for the studied measurements were narrower than the reported intra- and inter-observer variability among expert human sonographers for all the parameters under study. This means that these models could be deployed in clinical conditions, to alleviate time-consuming, repetitive tasks, and make fetal ultrasound more accessible in limited-resource environments.

Using Deep Learning to Detect the Presence and Location of Hemoperitoneum on the Focused Assessment with Sonography in Trauma (FAST) Examination in Adults

Abdominal ultrasonography has become an integral component of the evaluation of trauma patients. Internal hemorrhage can be rapidly diagnosed by finding free fluid with point-of-care ultrasound (POCUS) and expedite decisions to perform lifesaving interventions. However, the widespread clinical application of ultrasound is limited by the expertise required for image interpretation. This study aimed to develop a deep learning algorithm to identify the presence and location of hemoperitoneum on POCUS to assist novice clinicians in accurate interpretation of the Focused Assessment with Sonography in Trauma (FAST) exam. We analyzed right upper quadrant (RUQ) FAST exams obtained from 94 adult patients (44 confirmed hemoperitoneum) using the YoloV3 object detection algorithm. Exams were partitioned via fivefold stratified sampling for training, validation, and hold-out testing. We assessed each exam image-by-image using YoloV3 and determined hemoperitoneum presence for the exam using the detection with highest confidence score. We determined the detection threshold as the score that maximizes the geometric mean of sensitivity and specificity over the validation set. The algorithm had 95% sensitivity, 94% specificity, 95% accuracy, and 97% AUC over the test set, significantly outperforming three recent methods. The algorithm also exhibited strength in localization, while the detected box sizes varied with a 56% IOU averaged over positive cases. Image processing demonstrated only 57-ms latency, which is adequate for real-time use at the bedside. These results suggest that a deep learning algorithm can rapidly and accurately identify the presence and location of free fluid in the RUQ of the FAST exam in adult patients with hemoperitoneum.

The role of artificial intelligence in the future of urogynecology

Artificial intelligence (AI) in medicine is a rapidly growing field aimed at using machine learning models to improve health outcomes and patient experiences. Many new platforms have become accessible and therefore it seems inevitable that we consider how to implement them in our day-to-day practice. Currently, the specialty of urogynecology faces new challenges as the population grows, life expectancy increases, and quality of life expectation is much improved. As AI has a lot of potential to promote the discipline of urogynecology, we aim to explore its abilities and possible use in the future. Challenges and risks are associated with using AI, and a responsible use of such resources is required.

Automated detection and localization of pericardial effusion from point-of-care cardiac ultrasound examination

Focused Assessment with Sonography in Trauma (FAST) exam is the standard of care for pericardial and abdominal free fluid detection in emergency medicine. Despite its life saving potential, FAST is underutilized due to requiring clinicians with appropriate training and practice. To aid ultrasound interpretation, the role of artificial intelligence has been studied, while leaving room for improvement in localization information and computation time. The purpose of this study was to develop and test a deep learning approach to rapidly and accurately identify both the presence and location of pericardial effusion on point-of-care ultrasound (POCUS) exams. Each cardiac POCUS exam is analyzed image-by-image via the state-of-the-art YoloV3 algorithm and pericardial effusion presence is determined from the most confident detection. We evaluate our approach over a dataset of POCUS exams (cardiac component of FAST and ultrasound), comprising 37 cases with pericardial effusion and 39 negative controls. Our algorithm attains 92% specificity and 89% sensitivity in pericardial effusion identification, outperforming existing deep learning approaches, and localizes pericardial effusion by 51% Intersection Over Union with ground-truth annotations. Moreover, image processing demonstrates only 57 ms latency. Experimental results demonstrate the feasibility of rapid and accurate pericardial effusion detection from POCUS exams for physician overread.

Diagnosis of Pregnancy Complications Using Blind Ultrasound Sweeps Performed by Individuals Without Prior Formal Ultrasound Training

OBJECTIVE

To estimate the diagnostic accuracy of blind ultrasound sweeps performed with a low-cost, portable ultrasound system by individuals with no prior formal ultrasound training to diagnose common pregnancy complications.

METHODS

This is a single-center, prospective cohort study conducted from October 2020 to January 2022 among people with second- and third-trimester pregnancies. Nonspecialists with no prior formal ultrasound training underwent a brief training on a simple eight-step approach to performing a limited obstetric ultrasound examination that uses blind sweeps of a portable ultrasound probe based on external body landmarks. The sweeps were interpreted by five blinded maternal-fetal medicine subspecialists. Sensitivity, specificity, and positive and negative predictive values for blinded ultrasound sweep identification of pregnancy complications (fetal malpresentation, multiple gestations, placenta previa, and abnormal amniotic fluid volume) were compared with a reference standard ultrasonogram as the primary analysis. Kappa for agreement was also assessed.

RESULTS

Trainees performed 194 blinded ultrasound examinations on 168 unique pregnant people (248 fetuses) at a mean of 28±5.85 weeks of gestation for a total of 1,552 blinded sweep cine clips. There were 49 ultrasonograms with normal results (control group) and 145 ultrasonograms with abnormal results with known pregnancy complications. In this cohort, the sensitivity for detecting a prespecified pregnancy complication was 91.7% (95% CI 87.2-96.2%) overall, with the highest detection rate for multiple gestations (100%, 95% CI 100-100%) and noncephalic presentation (91.8%, 95% CI 86.4-97.3%). There was high negative predictive value for placenta previa (96.1%, 95% CI 93.5-98.8%) and abnormal amniotic fluid volume (89.5%, 95% CI 85.3-93.6%). There was also substantial to perfect mean agreement for these same outcomes (range 87-99.6% agreement, Cohen κ range 0.59-0.91, P<.001 for all).

CONCLUSION

Blind ultrasound sweeps of the gravid abdomen guided by an eight-step protocol using only external anatomic landmarks and performed by previously untrained operators with a low-cost, portable, battery-powered device had excellent sensitivity and specificity for high-risk pregnancy complications such as malpresentation, placenta previa, multiple gestations, and abnormal amniotic fluid volume, similar to results of a diagnostic ultrasound examination using a trained ultrasonographer and standard-of-care ultrasound machine. This approach has the potential to improve access to obstetric ultrasonography globally.

Understanding clinical outcome measures reported in HIV pregnancy studies involving antiretroviral-naive and antiretroviral-experienced women

HIV infection is a clinically significant public health disease and contributes to increased risk of maternal and fetal morbidity and mortality. HIV pregnancy studies use outcome measures as metrics to show how people with HIV feel, function, or survive. These endpoints are crucial for tracking the evolution of HIV illness over time, assessing the effectiveness of antiretroviral therapy (ART), and comparing outcomes across studies. Although the need for ideal outcome measures is widely acknowledged, selecting acceptable outcome measures for these HIV pregnancy studies can be challenging. We discuss the many outcome measures that have been implemented over time to assess HIV in pregnancy studies, their benefits, and drawbacks. Finally, we offer suggestions for improving the reporting of outcome measures in HIV in pregnancy studies. Medical professionals can best care for pregnant women living with HIV receiving ART by having a thorough understanding of these outcome metrics.

Development of a Machine Learning Model for Sonographic Assessment of Gestational Age

IMPORTANCE

Fetal ultrasonography is essential for confirmation of gestational age (GA), and accurate GA assessment is important for providing appropriate care throughout pregnancy and for identifying complications, including fetal growth disorders. Derivation of GA from manual fetal biometry measurements (ie, head, abdomen, and femur) is operator dependent and time-consuming.

OBJECTIVE

To develop artificial intelligence (AI) models to estimate GA with higher accuracy and reliability, leveraging standard biometry images and fly-to ultrasonography videos.

DESIGN, SETTING, AND PARTICIPANTS

To improve GA estimates, this diagnostic study used AI to interpret standard plane ultrasonography images and fly-to ultrasonography videos, which are 5- to 10-second videos that can be automatically recorded as part of the standard of care before the still image is captured. Three AI models were developed and validated: (1) an image model using standard plane images, (2) a video model using fly-to videos, and (3) an ensemble model (combining both image and video models). The models were trained and evaluated on data from the Fetal Age Machine Learning Initiative (FAMLI) cohort, which included participants from 2 study sites at Chapel Hill, North Carolina (US), and Lusaka, Zambia. Participants were eligible to be part of this study if they received routine antenatal care at 1 of these sites, were aged 18 years or older, had a viable intrauterine singleton pregnancy, and could provide written consent. They were not eligible if they had known uterine or fetal abnormality, or had any other conditions that would make participation unsafe or complicate interpretation. Data analysis was performed from January to July 2022.

MAIN OUTCOMES AND MEASURES

The primary analysis outcome for GA was the mean difference in absolute error between the GA model estimate and the clinical standard estimate, with the ground truth GA extrapolated from the initial GA estimated at an initial examination.

RESULTS

Of the total cohort of 3842 participants, data were calculated for a test set of 404 participants with a mean (SD) age of 28.8 (5.6) years at enrollment. All models were statistically superior to standard fetal biometry-based GA estimates derived from images captured by expert sonographers. The ensemble model had the lowest mean absolute error compared with the clinical standard fetal biometry (mean [SD] difference, -1.51 [3.96] days; 95% CI, -1.90 to -1.10 days). All 3 models outperformed standard biometry by a more substantial margin on fetuses that were predicted to be small for their GA.

CONCLUSIONS AND RELEVANCE

These findings suggest that AI models have the potential to empower trained operators to estimate GA with higher accuracy.

A mobile-optimized artificial intelligence system for gestational age and fetal malpresentation assessment

Background

Fetal ultrasound is an important component of antenatal care, but shortage of adequately trained healthcare workers has limited its adoption in low-to-middle-income countries. This study investigated the use of artificial intelligence for fetal ultrasound in under-resourced settings.

Methods

Blind sweep ultrasounds, consisting of six freehand ultrasound sweeps, were collected by sonographers in the USA and Zambia, and novice operators in Zambia. We developed artificial intelligence (AI) models that used blind sweeps to predict gestational age (GA) and fetal malpresentation. AI GA estimates and standard fetal biometry estimates were compared to a previously established ground truth, and evaluated for difference in absolute error. Fetal malpresentation (non-cephalic vs cephalic) was compared to sonographer assessment. On-device AI model run-times were benchmarked on Android mobile phones.

Results

Here we show that GA estimation accuracy of the AI model is non-inferior to standard fetal biometry estimates (error difference -1.4 ± 4.5 days, 95% CI -1.8, -0.9, n = 406). Non-inferiority is maintained when blind sweeps are acquired by novice operators performing only two of six sweep motion types. Fetal malpresentation AUC-ROC is 0.977 (95% CI, 0.949, 1.00, n = 613), sonographers and novices have similar AUC-ROC. Software run-times on mobile phones for both diagnostic models are less than 3 s after completion of a sweep.

Conclusions

The gestational age model is non-inferior to the clinical standard and the fetal malpresentation model has high AUC-ROCs across operators and devices. Our AI models are able to run on-device, without internet connectivity, and provide feedback scores to assist in upleveling the capabilities of lightly trained ultrasound operators in low resource settings.

The Basics of Machine Learning

The Basics of Machine LearningWhen a person is pregnant, a key question is how to establish the "date" of the pregnancy. Classically, the date was based on the last menstrual period (LMP). For the past 3 decades or more, in high-resource countries, this has been done using "hospital-grade" ultrasound machines, with testing performed by trained sonographers. In many parts of the world, neither the machines nor the trained sonographers are accessible. In an article published in NEJM Evidence, Pokaprakarn et al.1 asked whether a low-cost handheld ultrasound device combined with artificial intelligence (AI) could substitute for the expensive machines and trained sonographers.

Leveraging Artificial Intelligence to Improve Pregnancy Dating in Low-Resource Settings

For practicing obstetricians and pediatricians, accurate gestational age (GA) is a cornerstone of quality care during pregnancy and the newborn period. GA determines prenatal management decisions, such as eligibility for antenatal glucocorticoids if preterm delivery is anticipated or timing of delivery in complicated pregnancies to avoid stillbirth or maternal morbidity. Knowledge of GA for women in labor also allows for transfer to facilities capable of handling preterm infants and guides postnatal resuscitation practices and specialized newborn care.