Holdout Dataset Research Articles

ValveVision AI - a multimodal language model for qualitative reporting of the aortic valve in echocardiography

Abstract Background The precise assessment of the aortic valve via echocardiography is critical for early detection and management of aortic valve diseases. Until recently, previous studies have examined machine learning models to estimate individual measurements and severity of aortic stenosis (AS) from echocardiographic images. These image processing algorithms, while precise in their narrow focus, fall short in mirroring the holistic and interconnected clinical judgment typical of human echocardiographers in producing a qualitative report. Large language models (LLMs), particularly image-to-text multi-modal LLMs, are a fundamental advance in the field of deep learning with implications for a host of applications in medical imaging. They promise to encapsulate not just discrete data points typical in traditional machine learning, but also the complex contextual interrelations in clinical diagnosis. Methods In this study, a large-scale heterogeneous database of echocardiographic images containing over 90,681 studies with textual descriptors of the aortic valve was used to train a single, image-to-text multimodal LLM called ValveVision AI. The ground truth textual summaries were drafted by level III echocardiographers in a clinical setting between 2015-2020. BLEU and ROUGE score was calculated. The models were retrospectively assessed on a holdout dataset. Reviewing physicians compared the generated summary to the ground truth and binarily agreed or rejected it. Receiver Operating Characteristics (ROC) for distinct pathologies were also assessed (Figure I). Results ValveVision AI performed with a BLEU score of 0.45 and a ROUGE score of 0.49. The performance of the model in reporting on classification of moderate/severe vs none/mild AS in concordance with the validation protocol described above achieved a specificity of 91.98% and a sensitivity of 83.89%, along with more precise qualitative description. Qualitatively, the model exhibited the capability of zero-shot learning in certain instances, however, this result remains an area of exploration. Conclusion This study represents to our knowledge, the first attempt at an image-representation to text-tokenizer deep learning model architecture to mimic the thought and subtlety of echocardiographic qualitative analysis of the aortic valve. The results suggest that this multimodal LLM has sufficient accuracy to create a preliminary textual summary of the aortic valve that, if paired with a point of care ultrasound (POCUS) device in a primary care setting, may facilitate case triage, increase efficiency, and determine a more precise care pathway for patients.

Comorbidity neural network model for personalised recommendations in cardiometabolic risk factor management: towards early precision medicine in routine clinical practice

Abstract Background and Aims Hypertension and hypercholesterolaemia often develop in tandem to confer a risk of cardiovascular disease that may be more than additive. The present study explored early comorbid therapeutic intervention, i.e., beyond individual disease silos, while minimising data bias to provide clinical decision support tools that are portable across clinics. Methods A neural network was created to reproduce UK national clinical guidelines with an accuracy higher than 95%. Relying on real-world evidence from the UK primary care Clinical Practice Research Datalink electronic health record resource, a transfer learning methodology was then applied to achieve recommended risk factor goal attainment after two years. Shapley values derived from game theory were used to define similar digital twin cohorts, and to evaluate individualised recommendations that deviated from national clinical guidelines by rejecting a no-benefit hypothesis retrospectively in a classical proportion test. The methodology was ported to an independent hold-out dataset with anonymized records from six primary care clinics in the South of England. Results The adult patient population (n=469,496, mean+SD age 61+11yr; M 56% vs F 44%; primary 84% vs secondary prevention 16%) included 101,255 diagnosed for hypertension and 112,180 for cholesterol management. While a majority of the 52,504 therapeutic decisions in a test set agree with single morbidity guidelines, for 4639 deviations from the guidance, a significant minority was identified for whom alternative therapy was recommended. Rejecting a no-clinical benefit hypothesis at the 95% confidence level in a retrospective analysis, 1367 decisions showed statistically significant better outcomes over guidelines, for example by escalating anti-hypertensive therapies for patients receiving statin therapy (1085) or by complementing anti-hypertensive therapy with a statin (158). Similar findings were obtained when tested in a hold-out dataset with 13,725 decisions for 6531 patients. Conclusions The results support the contention that machine learning can identify patient subgroups for whom comorbid intervention deviating from clinical guidelines may be appropriate. A novel aspect of the study is the demonstration of clinically portable and actionable application of machine learning in the context of comorbid cardiometabolic diseases. The results represent practical steps towards personalised precision medicine for cardiovascular comorbidity in routine clinical practice.

Prediction model for major bleeding in anticoagulated patients with cancer-associated venous thromboembolism using machine learning and natural language processing.

We developed a predictive model to assess the risk of major bleeding (MB) within 6 months of primary venous thromboembolism (VTE) in cancer patients receiving anticoagulant treatment. We also sought to describe the prevalence and incidence of VTE in cancer patients, and to describe clinical characteristics at baseline and bleeding events during follow-up in patients receiving anticoagulants. This observational, retrospective, and multicenter study used natural language processing and machine learning (ML), to analyze unstructured clinical data from electronic health records from nine Spanish hospitals between 2014 and 2018. All adult cancer patients with VTE receiving anticoagulants were included. Both clinically- and ML-driven feature selection was performed to identify MB predictors. Logistic regression (LR), decision tree (DT), and random forest (RF) algorithms were used to train predictive models, which were validated in a hold-out dataset and compared to the previously developed CAT-BLEED score. Of the 2,893,108 cancer patients screened, in-hospital VTE prevalence was 5.8% and the annual incidence ranged from 2.7 to 3.9%. We identified 21,227 patients with active cancer and VTE receiving anticoagulants (53.9% men, median age of 70years). MB events after VTE diagnosis occurred in 10.9% of patients within the first six months. MB predictors included: hemoglobin, metastasis, age, platelets, leukocytes, and serum creatinine. The LR, DT, and RF models had AUC-ROC (95% confidence interval) values of 0.60 (0.55, 0.65), 0.60 (0.55, 0.65), and 0.61 (0.56, 0.66), respectively. These models outperformed the CAT-BLEED score with values of 0.53 (0.48, 0.59). Our study shows encouraging results in identifying anticoagulated patients with cancer-associated VTE who are at high risk of MB.

Topology-Driven Discovery of Transmembrane Protein S-Palmitoylation.

Protein S-palmitoylation is a reversible lipophilic posttranslational modification regulating a diverse number of signaling pathways. Within transmembrane proteins (TMPs), S-palmitoylation is implicated in conditions from inflammatory disorders to respiratory viral infections. Many small-scale experiments have observed S-palmitoylation at juxtamembrane Cys residues. However, most large-scale S-palmitoyl discovery efforts rely on trypsin-based proteomics within which hydrophobic juxtamembrane regions are likely underrepresented. Machine learning- by virtue of its freedom from experimental constraints - is particularly well suited to address this discovery gap surrounding TMP S-palmitoylation. Utilizing a UniProt-derived feature set, a gradient boosted machine learning tool (TopoPalmTree) was constructed and applied to a holdout dataset of viral S-palmitoylated proteins. Upon application to the mouse TMP proteome, 1591 putative S-palmitoyl sites (i.e. not listed in SwissPalm or UniProt) were identified. Two lung-expressed S-palmitoyl candidates (synaptobrevin Vamp5 and water channel Aquaporin-5) were experimentally assessed. Finally, TopoPalmTree was used for rational design of an S-palmitoyl site on KDEL-Receptor 2. This readily interpretable model aligns the innumerable small-scale experiments observing juxtamembrane S-palmitoylation into a proteomic tool for TMP S-palmitoyl discovery and design, thus facilitating future investigations of this important modification.

Major depletion of insulin sensitivity-associated taxa in the gut microbiome of persons living with HIV controlled by antiretroviral drugs

BackgroundPersons living with HIV (PWH) harbor an altered gut microbiome (higher abundance of Prevotella and lower abundance of Bacillota and Ruminococcus lineages) compared to non-infected individuals. Some of these alterations are linked to sexual preference and others to the HIV infection. The relationship between these lineages and metabolic alterations, often present in aging PWH, has been poorly investigated.MethodsIn this study, we compared fecal metagenomes of 25 antiretroviral-treatment (ART)-controlled PWH to three independent control groups of 25 non-infected matched individuals by means of univariate analyses and machine learning methods. Moreover, we used two external datasets to validate predictive models of PWH classification. Next, we searched for associations between clinical and biological metabolic parameters with taxonomic and functional microbiome profiles. Finally, we compare the gut microbiome in 7 PWH after a 17-week ART switch to raltegravir/maraviroc.ResultsThree major enterotypes (Prevotella, Bacteroides and Ruminococcaceae) were present in all groups. The first Prevotella enterotype was enriched in PWH, with several of characteristic lineages associated with poor metabolic profiles (low HDL and adiponectin, high insulin resistance (HOMA-IR)). Conversely butyrate-producing lineages were markedly depleted in PWH independently of sexual preference and were associated with a better metabolic profile (higher HDL and adiponectin and lower HOMA-IR). Accordingly with the worst metabolic status of PWH, butyrate production and amino-acid degradation modules were associated with high HDL and adiponectin and low HOMA-IR. Random Forest models trained to classify PWH vs. control on taxonomic abundances displayed high generalization performance on two external holdout datasets (ROC AUC of 80–82%). Finally, no significant alterations in microbiome composition were observed after switching to raltegravir/maraviroc.ConclusionHigh resolution metagenomic analyses revealed major differences in the gut microbiome of ART-controlled PWH when compared with three independent matched cohorts of controls. The observed marked insulin resistance could result both from enrichment in Prevotella lineages, and from the depletion in species producing butyrate and involved into amino-acid degradation, which depletion is linked with the HIV infection.

Enhancing the reliability of deep learning-based head and neck tumour segmentation using uncertainty estimation with multi-modal images

Objective. Deep learning shows promise in autosegmentation of head and neck cancer (HNC) primary tumours (GTV-T) and nodal metastases (GTV-N). However, errors such as including non-tumour regions or missing nodal metastases still occur. Conventional methods often make overconfident predictions, compromising reliability. Incorporating uncertainty estimation, which provides calibrated confidence intervals can address this issue. Our aim was to investigate the efficacy of various uncertainty estimation methods in improving segmentation reliability. We evaluated their confidence levels in voxel predictions and ability to reveal potential segmentation errors. Approach. We retrospectively collected data from 567 HNC patients with diverse cancer sites and multi-modality images (CT, PET, T1-, and T2-weighted MRI) along with their clinical GTV-T/N delineations. Using the nnUNet 3D segmentation pipeline, we compared seven uncertainty estimation methods, evaluating them based on segmentation accuracy (Dice similarity coefficient, DSC), confidence calibration (Expected Calibration Error, ECE), and their ability to reveal segmentation errors (Uncertainty-Error overlap using DSC, UE-DSC). Main results. Evaluated on the hold-out test dataset (n = 97), the median DSC scores for GTV-T and GTV-N segmentation across all uncertainty estimation methods had a narrow range, from 0.73 to 0.76 and 0.78 to 0.80, respectively. In contrast, the median ECE exhibited a wider range, from 0.30 to 0.12 for GTV-T and 0.25 to 0.09 for GTV-N. Similarly, the median UE-DSC also ranged broadly, from 0.21 to 0.38 for GTV-T and 0.22 to 0.36 for GTV-N. A probabilistic network—PhiSeg method consistently demonstrated the best performance in terms of ECE and UE-DSC. Significance. Our study highlights the importance of uncertainty estimation in enhancing the reliability of deep learning for autosegmentation of HNC GTV. The results show that while segmentation accuracy can be similar across methods, their reliability, measured by calibration error and uncertainty-error overlap, varies significantly. Used with visualisation maps, these methods may effectively pinpoint uncertainties and potential errors at the voxel level.

A Deep Transfer Learning Approach for Sleep Stage Classification and Sleep Apnea Detection Using Wrist-Worn Consumer Sleep Technologies.

Obstructive sleep apnea (OSA) is a common, underdiagnosed sleep-related breathing disorder with serious health implications Objective - We propose a deep transfer learning approach for sleep stage classification and sleep apnea (SA) detection using wrist-worn consumer sleep technologies (CST). Methods - Our model is based on a deep convolutional neural network (DNN) utilizing accelerometers and photo-plethysmography signals from nocturnal recordings. The DNN was trained and tested on internal datasets that include raw data from clinical and wrist-worn devices; external validation was performed on a hold-out test dataset containing raw data from a wrist-worn CST. Results - Training on clinical data improves performance significantly, and feature enrichment through a sleep stage stream gives only minor improvements. Raw data input outperforms feature-based input in CST datasets. The system generalizes well but performs slightly worse on wearable device data compared to clinical data. However, it excels in detecting events during REM sleep and is associated with arousal and oxygen desaturation. We found; cases that were significantly underestimated were characterized by fewer of such event associations. Conclusion - This study showcases the potential of using CSTs as alternate screening solution for undiagnosed cases of OSA. Significance - This work is significant for its development of a deep transfer learning approach using wrist-worn consumer sleep technologies, offering comprehensive validation for data utilization, and learning techniques, ultimately improving sleep apnea detection across diverse devices.

Adaptive fine-tuning based transfer learning for the identification of MGMT promoter methylation status.

Background.Glioblastoma Multiforme (GBM) is an aggressive form of malignant brain tumor with a generally poor prognosis.O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation has been shown to be a predictive bio-marker for resistance to treatment of GBM, but it is invasive and time-consuming to determine methylation status. There has been effort to predict the MGMT methylation status through analyzing MRI scans using machine learning, which only requires pre-operative scans that are already part of standard-of-care for GBM patients.Purpose.To improve the performance of conventional transfer learning in the identification of MGMT promoter methylation status, we developed a 3D SpotTune network with adaptive fine-tuning capability. Using the pretrained weights of MedicalNet with the SpotTune network, we compared its performance with a randomly initialized network for different combinations of MR modalities.Methods.Using a ResNet50 as the base network, three categories of networks are created: (1) A 3D SpotTune network to process volumetric MR images, (2) a network with randomly initialized weights, and (3) a network pre-trained on MedicalNet. These three networks are trained and evaluated using a public GBM dataset provided by the University of Pennsylvania. The MRI scans from 240 patients are used, with 11 different modalities corresponding to a set of perfusion, diffusion, and structural scans. The performance is evaluated using 5-fold cross validation with a hold-out testing dataset.Results.The SpotTune network showed better performance than the randomly initialized network. The best performing SpotTune model achieved an area under the Receiver Operating Characteristic curve (AUC), average precision of the precision-recall curve (AP), sensitivity, and specificity values of 0.6604, 0.6179, 0.6667, and 0.6061 respectively.Conclusions.SpotTune enables transfer learning to be adaptive to individual patients, resulting in improved performance in predicting MGMT promoter methylation status in GBM using equivalent MRI modalities as compared to a randomly initialized network.

Prognostic value of platelet levels in patients with aneurysmal Subarachnoid Hemorrhage

Pathophysiological processes following aneurysmal subarachnoid hemorrhage (aSAH) include upregulated underlying systemic inflammation, which is reflected by changes in different peripheral blood cells and their sub-populations. As inflammation is a crucial process that contributes to post-aSAH complications and clincal outcome, blood cell numbers and ratios in systemic circulation may predict the outcome and provide rapid and easy to quantify point of care biomarkers for these critically ill patients. To identify blood-derived cellular inflammatory parameters which allow a precise prediction of patient outcome after aSAH. In this single-center retrospective study, 19 whole blood-derived cellular inflammatory markers and clinical and demographic parameters for 101 aSAH patients were recorded within 24 h after aSAH. Clinical outcome was quantified with modified Rankin scale (mRS) on discharge. Proportional odds logistic regression (POLR) was used to model the patient outcome as the function of clinical parameters and inflammatory markers. The results were validated on a separate hold-out dataset (220 patients). The on-admission platelet count, mean platelet volume (MPV) and mean platelet volume to platelet ratio (MPR) were found to be significant and predictive of patient outcome on discharge. Mean platelet volume (MPV) and mean platelet volume to platelet ratio (MPR) predicted clinical outcome and may serve as easy to quantify point of care biomarker. The findings are potentially relevant for the management of aSAH.

Novel Domain Knowledge-Encoding Algorithm Enables Label-Efficient Deep Learning for Cardiac CT Segmentation to Guide Atrial Fibrillation Treatment in a Pilot Dataset.

Background: Segmenting computed tomography (CT) is crucial in various clinical applications, such as tailoring personalized cardiac ablation for managing cardiac arrhythmias. Automating segmentation through machine learning (ML) is hindered by the necessity for large, labeled training data, which can be challenging to obtain. This article proposes a novel approach for automated, robust labeling using domain knowledge to achieve high-performance segmentation by ML from a small training set. The approach, the domain knowledge-encoding (DOKEN) algorithm, reduces the reliance on large training datasets by encoding cardiac geometry while automatically labeling the training set. The method was validated in a hold-out dataset of CT results from an atrial fibrillation (AF) ablation study. Methods: The DOKEN algorithm parses left atrial (LA) structures, extracts "anatomical knowledge" by leveraging digital LA models (available publicly), and then applies this knowledge to achieve high ML segmentation performance with a small number of training samples. The DOKEN-labeled training set was used to train a nnU-Net deep neural network (DNN) model for segmenting cardiac CT in N = 20 patients. Subsequently, the method was tested in a hold-out set with N = 100 patients (five times larger than training set) who underwent AF ablation. Results: The DOKEN algorithm integrated with the nn-Unet model achieved high segmentation performance with few training samples, with a training to test ratio of 1:5. The Dice score of the DOKEN-enhanced model was 96.7% (IQR: 95.3% to 97.7%), with a median error in surface distance of boundaries of 1.51 mm (IQR: 0.72 to 3.12) and a mean centroid-boundary distance of 1.16 mm (95% CI: -4.57 to 6.89), similar to expert results (r = 0.99; p < 0.001). In digital hearts, the novel DOKEN approach segmented the LA structures with a mean difference for the centroid-boundary distances of -0.27 mm (95% CI: -3.87 to 3.33; r = 0.99; p < 0.0001). Conclusions: The proposed novel domain knowledge-encoding algorithm was able to perform the segmentation of six substructures of the LA, reducing the need for large training data sets. The combination of domain knowledge encoding and a machine learning approach could reduce the dependence of ML on large training datasets and could potentially be applied to AF ablation procedures and extended in the future to other imaging, 3D printing, and data science applications.

Contrastive Self-supervised Learning for Neurodegenerative Disorder Classification.

Neurodegenerative diseases such as Alzheimer's disease (AD) or frontotemporal lobar degeneration (FTLD) involve specific loss of brain volume, detectable in vivo using T1-weighted MRI scans. Supervised machine learning approaches classifying neurodegenerative diseases require diagnostic-labels for each sample. However, it can be difficult to obtain expert labels for a large amount of data. Self-supervised learning (SSL) offers an alternative for training machine learning models without data-labels. We investigated if the SSL models can applied to distinguish between different neurodegenerative disorders in an interpretable manner. Our method comprises a feature extractor and a downstream classification head. A deep convolutional neural network trained in a contrastive self-supervised way serves as the feature extractor, learning latent representation, while the classifier head is a single-layer perceptron. We used N=2694 T1-weighted MRI scans from four data cohorts: two ADNI datasets, AIBL and FTLDNI, including cognitively normal controls (CN), cases with prodromal and clinical AD, as well as FTLD cases differentiated into its sub-types. Our results showed that the feature extractor trained in a self-supervised way provides generalizable and robust representations for the downstream classification. For AD vs. CN, our model achieves 82% balanced accuracy on the test subset and 80% on an independent holdout dataset. Similarly, the Behavioral variant of frontotemporal dementia (BV) vs. CN model attains an 88% balanced accuracy on the test subset. The average feature attribution heatmaps obtained by the Integrated Gradient method highlighted hallmark regions, i.e., temporal gray matter atrophy for AD, and insular atrophy for BV. In conclusion, our models perform comparably to state-of-the-art supervised deep learning approaches. This suggests that the SSL methodology can successfully make use of unannotated neuroimaging datasets as training data while remaining robust and interpretable.

O-219 Explainable artificial intelligence (XAI) to determine the follicle sizes on trigger day that maximize mature oocyte yield

Abstract Study question Which follicle sizes on trigger day (TD) are key to obtaining mature oocytes, embryos, and blastocysts, thereby enhancing live birth rates? Summary answer Follicles sized 13-18mm are most likely to yield mature oocytes. Live birth rate (LBR) was increased with a greater proportion of follicles in this range. What is known already At the time of oocyte retrieval, follicles that are either too small or too large are less likely to yield mature oocytes. However, there is only limited data examining which follicle sizes on TD administration are optimally sized to yield mature oocytes, nor any data examining whether live birth rates are impacted. Explainable artificial intelligence (XAI) offers a valuable opportunity to leverage clinical data to determine the range of follicle sizes that are most likely to yield mature oocytes and lead to improved downstream outcomes. Study design, size, duration A retrospective study incorporating 19,082 patients undergoing their first IVF/ICSI cycle (2005-2023) across 11 European clinics. Follicle sizes on TD were related to oocyte maturity (14,140 patients), fertilization (17,822 patients), high-quality blastocyst development (17,488 patients), and pregnancy outcomes (12,672 patients). Secondary analyses included stratification by age (&amp;gt;35 years: 4,717 patients; ≤35 years: 5,707), and suppressant protocol (GnRH agonist: n = 5,420 patients; GnRH antagonist: n = 3,982). Participants/materials, setting, methods An ensemble-based XAI model, developed through ‘internal-external validation’, was trained, Bayesian optimized, and independently iteratively tested on hold-out clinic datasets. The model determined follicle sizes contributing most to successful laboratory outcomes by harnessing permutation importance and ‘SHAP’ values. Model performance was evaluated and optimized for mean absolute error (MAE). The top 50% contiguous cohort of influential follicle sizes on TD were identified, and further validated through analysis of follicle sizes on preceding days. Main results and the role of chance XAI analysis revealed that follicles sized 13-18mm on TD were most influential in yielding mature oocytes, especially those sized 15-18mm (MAE: 3.60 ±0.35). Predictive follicle sizes from scans 1-2 days preceding TD administration were consistent with expected mean follicle growth rates. For high-quality blastocysts, follicles sized 14-20mm were contributory, narrowing to 15-18mm in the subset of patients where oocytes had been confirmed to be mature. In hCG-triggered cycles, follicles sized 12-19mm were most likely to yield mature oocytes in GnRH antagonist cycles, whereas follicles sized 14-20mm were most contributory in GnRH agonist cycles. In patients ≤35 years, follicles 13-18mm were most contributory, whereas a wider range of follicles (11-20mm) were contributory in patients &amp;gt;35 years. In line with current practice, having three lead follicles ≥17mm on TD was associated with a median improvement in mature oocyte yield of 10% versus those that did not (p &amp;lt; 0.0001). By comparison, maximizing the proportion of follicles on TD within the optimal size range could better improve mature oocyte yield, for example, by 42% when at least 70% of follicles were sized 15-18mm (p &amp;lt; 0.0001). Likewise, LBR improved as the proportion of follicles sized 13-18mm was increased (&amp;lt;10%: 23.3%; 10-20%: 25.7%; 20-30%: 28.8%; 30-40%: 31.6%). Limitations, reasons for caution These data could help guide when to administer the trigger to optimize clinical outcomes, however prospective assessment of different TD strategies based on the proportion of follicles within the optimal follicle size ranges identified in comparison to current practice based on lead follicle size is required prior to clinical adoption. Wider implications of the findings These findings provide a data-driven approach to personalizing ovarian stimulation in IVF. The findings generated by applying an XAI model could assist clinicians in optimizing TD timing, potentially enhancing clinical outcomes by moving beyond the conventional focus on lead follicle size. Trial registration number not appicable

LGG-15. DEEP LEARNING ENABLES LONGITUDINAL RISK PREDICTION FOR PEDIATRIC LOW-GRADE GLIOMAS AFTER SURGERY

Abstract BACKGROUND Disease progression is challenging to predict following surgery for pediatric low-grade glioma (pLGG), and early progression indicators on MRI surveillance imaging would help guide management. Longitudinal surveillance imaging data may capture subtle temporal tumor changes and patterns that could inform recurrence risk but are difficult to synthesize clinically. We applied deep, self-supervised learning to longitudinal surveillance MRI to predict short-interval event-free-survival (EFS) risk from time-of-scan. METHODS We retrospectively collected data from 278 patients (2,174 scans) who underwent surgical resection for pLGG at Dana-Farber Cancer Institute/Boston Children’s Hospital (DFCI/BCH) from 1990-2022. A deep-learning model was pretrained with T2-FLAIR sequences to determine the chronological scan order for each patient (temporal pretraining). The model was fine-tuned to predict 1-year EFS from point-of-scan using a subset of scans from DFCI/BCH and then validated on a hold-out dataset (511 scans, 67 patients). Separately, the model was fine-tuned on an external dataset from the Children’s Brain Tumor Network (CBTN) and validated on a hold-out set (204 scans, 61 patients). Results were compared to a model trained without temporal pretraining. RESULTS Median follow-up and interval between scans were 73 and 6.3 months for DFCI/BCH and 20 and 5.8 months for CBTN. Temporal pretraining improved performance for 1-year EFS, for DFCI/BCH (AUC: 0.83 [0.71-0.90] vs. 0.78 [0.62-0.91]; F1-score: 0.80 vs. 0.67), and CBTN hold-out sets (AUC: 0.75 [0.58-0.90] vs. 0.53 [0.35-0.70], p=0.01; F1-score: 0.65 vs. 0.55). Increasing the number of longitudinal scans available to the model incrementally improved performance up to 7 scans/patient for DFCI/BCH and 10 scans/patient for CBTN. CONCLUSIONS Deep-learning with temporal pretraining improves the predictive analysis of tumor growth patterns in postoperative pLGG surveillance MRI, enabling accurate short-interval EFS prediction. This model could help inform patients and clinicians regarding surveillance regimen and initiation of adjuvant therapy.

DentalSegmentator: Robust open source deep learning-based CT and CBCT image segmentation

ObjectivesSegmentation of anatomical structures on dento-maxillo-facial (DMF) computed tomography (CT) or cone beam computed tomography (CBCT) scans is increasingly needed in digital dentistry. The main aim of this research was to propose and evaluate a novel open source tool called DentalSegmentator for fully automatic segmentation of five anatomical structures on DMF CT and CBCT scans: maxilla/upper skull, mandible, upper teeth, lower teeth, and the mandibular canal. MethodsA retrospective sample of 470 CT and CBCT scans was used as a training/validation set. The performance and generalizability of the tool was evaluated by comparing segmentations provided by experts and automatic segmentations in two hold-out test datasets: an internal dataset of 133 CT and CBCT scans acquired before orthognathic surgery and an external dataset of 123 CBCT scans randomly sampled from routine examinations in 5 institutions. ResultsThe mean overall results in the internal test dataset (n = 133) were a Dice similarity coefficient (DSC) of 92.2 ± 6.3 % and a normalised surface distance (NSD) of 98.2 ± 2.2 %. The mean overall results on the external test dataset (n = 123) were a DSC of 94.2 ± 7.4 % and a NSD of 98.4 ± 3.6 %. ConclusionsThe results obtained from this highly diverse dataset demonstrate that this tool can provide fully automatic and robust multiclass segmentation for DMF CT and CBCT scans. To encourage the clinical deployment of DentalSegmentator, the pre-trained nnU-Net model has been made publicly available along with an extension for the 3D Slicer software. Clinical significanceDentalSegmentator open source 3D Slicer extension provides a free, robust, and easy-to-use approach to obtaining patient-specific three-dimensional models from CT and CBCT scans. These models serve various purposes in a digital dentistry workflow, such as visualization, treatment planning, intervention, and follow-up.

Pulmonary Fibrosis Diagnosis and Disease Progression Detected Via Hair Metabolome Analysis.

Fibrotic interstitial lung disease is often identified late due to non-specific symptoms, inadequate access to specialist care, and clinical unawareness precluding proper and timely treatment. Biopsy histological analysis is definitive but rarely performed due to its invasiveness. Diagnosis typically relies on high-resolution computed tomography, while disease progression is evaluated via frequent pulmonary function testing. This study tested the hypothesis that pulmonary fibrosis diagnosis and progression could be non-invasively and accurately evaluated from the hair metabolome, with the longer-term goal to minimize patient discomfort. Hair specimens collected from pulmonary fibrosis patients (n = 56) and healthy subjects (n = 14) were processed for metabolite extraction using 2DLC/MS-MS, and data were analyzed via machine learning. Metabolomic data were used to train machine learning classification models tuned via a rigorous combination of cross validation, feature selection, and testing with a hold-out dataset to evaluate classifications of diseased vs. healthy subjects and stable vs. progressed disease. Prediction of pulmonary fibrosis vs. healthy achieved AUROCTRAIN = 0.888 (0.794-0.982) and AUROCTEST = 0.908, while prediction of stable vs. progressed disease achieved AUROCTRAIN = 0.833 (0.784 - 0.882) and AUROCTEST = 0. 799. Top metabolites for diagnosis included ornithine, 4-(methylnitrosamino)-1-3-pyridyl-N-oxide-1-butanol, Thr-Phe, desthiobiotin, and proline. Top metabolites for progression included azelaic acid, Thr-Phe, Ala-Tyr, indoleacetyl glutamic acid, and cytidine. This study provides novel evidence that pulmonary fibrosis diagnosis and progression may in principle be evaluated from the hair metabolome. Longer term, this approach may facilitate non-invasive and accurate detection and monitoring of fibrotic lung diseases.

Epistatic Features and Machine Learning Improve Alzheimer's Disease Risk Prediction Over Polygenic Risk Scores.

Polygenic risk scores (PRS) are linear combinations of genetic markers weighted by effect size that are commonly used to predict disease risk. For complex heritable diseases such as late-onset Alzheimer's disease (LOAD), PRS models fail to capture much of the heritability. Additionally, PRS models are highly dependent on the population structure of the data on which effect sizes are assessed and have poor generalizability to new data. The goal of this study is to construct a paragenic risk score that, in addition to single genetic marker data used in PRS, incorporates epistatic interaction features and machine learning methods to predict risk for LOAD. We construct a new state-of-the-art genetic model for risk of Alzheimer's disease. Our approach innovates over PRS models in two ways: First, by directly incorporating epistatic interactions between SNP loci using an evolutionary algorithm guided by shared pathway information; and second, by estimating risk via an ensemble of non-linear machine learning models rather than a single linear model. We compare the paragenic model to several PRS models from the literature trained on the same dataset. The paragenic model is significantly more accurate than the PRS models under 10-fold cross-validation, obtaining an AUC of 83% and near-clinically significant matched sensitivity/specificity of 75%. It remains significantly more accurate when evaluated on an independent holdout dataset and maintains accuracy within APOE genotype strata. Paragenic models show potential for improving disease risk prediction for complex heritable diseases such as LOAD over PRS models.

Uncertainty estimation using a 3D probabilistic U-Net for segmentation with small radiotherapy clinical trial datasets

Background and objectivesBio-medical image segmentation models typically attempt to predict one segmentation that resembles a ground-truth structure as closely as possible. However, as medical images are not perfect representations of anatomy, obtaining this ground truth is not possible. A surrogate commonly used is to have multiple expert observers define the same structure for a dataset. When multiple observers define the same structure on the same image there can be significant differences depending on the structure, image quality/modality and the region being defined. It is often desirable to estimate this type of aleatoric uncertainty in a segmentation model to help understand the region in which the true structure is likely to be positioned. Furthermore, obtaining these datasets is resource intensive so training such models using limited data may be required. With a small dataset size, differing patient anatomy is likely not well represented causing epistemic uncertainty which should also be estimated so it can be determined for which cases the model is effective or not. MethodsWe use a 3D probabilistic U-Net to train a model from which several segmentations can be sampled to estimate the range of uncertainty seen between multiple observers. To ensure that regions where observers disagree most are emphasised in model training, we expand the Generalised Evidence Lower Bound (ELBO) with a Constrained Optimisation (GECO) loss function with an additional contour loss term to give attention to this region. Ensemble and Monte-Carlo dropout (MCDO) uncertainty quantification methods are used during inference to estimate model confidence on an unseen case. We apply our methodology to two radiotherapy clinical trial datasets, a gastric cancer trial (TOPGEAR, TROG 08.08) and a post-prostatectomy prostate cancer trial (RAVES, TROG 08.03). Each dataset contains only 10 cases each for model development to segment the clinical target volume (CTV) which was defined by multiple observers on each case. An additional 50 cases are available as a hold-out dataset for each trial which had only one observer define the CTV structure on each case. Up to 50 samples were generated using the probabilistic model for each case in the hold-out dataset. To assess performance, each manually defined structure was matched to the closest matching sampled segmentation based on commonly used metrics. ResultsThe TOPGEAR CTV model achieved a Dice Similarity Coefficient (DSC) and Surface DSC (sDSC) of 0.7 and 0.43 respectively with the RAVES model achieving 0.75 and 0.71 respectively. Segmentation quality across cases in the hold-out datasets was variable however both the ensemble and MCDO uncertainty estimation approaches were able to accurately estimate model confidence with a p-value < 0.001 for both TOPGEAR and RAVES when comparing the DSC using the Pearson correlation coefficient. ConclusionsWe demonstrated that training auto-segmentation models which can estimate aleatoric and epistemic uncertainty using limited datasets is possible. Having the model estimate prediction confidence is important to understand for which unseen cases a model is likely to be useful.

A digital pathology AI model to predict immune-oncology biomarker status and pembrolizumab response in a real-world cohort of patients with colorectal and breast cancer.

e13617 Background: Deep learning (DL) algorithms have garnered significant attention for prediction of biomarker status in digitized tissue slides. Here, we trained and validated a digital pathology algorithm to predict microsatellite instability (MSI)/mismatch repair deficiency (MMRd) in colorectal cancer (CRC) and PD-L1 status in breast cancer (BC) using a large real-world patient cohort. Methods: H&amp;E and immunohistochemistry (IHC) slides of CRC and BC from Caris Life Sciences (Phoenix, AZ) with known biomarker status were digitally scanned (CRC: N=27,494, 6.6% MSI-high by next-generation sequencing [NGS]; N=25,788, 6.4% MMRd by IHC; BC: N=16,069, 15.8% PD-L1+ by IHC). A transformer-based DL model (Wagner, Sophia J., et al., 2023) was trained to predict biomarker status, with 5-fold cross-validation. Accuracy, F1 score, area under the curve (AUC), and other metrics were calculated comparing the model to NGS/IHC “truth”. Risk stratification was performed on holdout datasets of pembrolizumab-treated CRC (N=422-454) and triple negative breast cancer (TNBC; N=513) pts. Post-treatment overall survival (OS) was calculated from start of pembrolizumab to last contact, per insurance claims data. Hazard ratios (HR) and 95% confidence intervals (CI) were calculated by Cox-Proportional hazards and compared between model and “truth”. Results: When applied to holdout datasets for biomarker prediction, the validated model achieved AUC of 0.78-0.95. In CRC and BC, performance was better among primary specimens than metastatic. For MSI-H/MMRd risk stratification in pembrolizumab-treated CRC, HR for OS were comparable between model and “truth” for primary and metastatic cohorts. When applied to the TNBC holdout dataset, the model performed better than “truth” for PD-L1 risk stratification in pembrolizumab-treated pts. Conclusions: A digital pathology DL model predicts biomarker status with high accuracy and is non-inferior to NGS/IHC for risk stratification of pembrolizumab-treated CRC and TNBC pts. These results support further investigation of AI algorithms developed from large datasets as possible rapid screening tools for biomarker detection in molecular testing programs. [Table: see text]

Design of a novel deep network model for spinal cord injury prediction

Degenerative cervical myelopathy must be diagnosed with magnetic resonance imaging (MRI) which predicts spinal cord injury (SCI). The growing volume of medical imaging data can be managed by deep learning models, which provide a preliminary interpretation of images taken in basic care settings. Our main goal was to create a deep-learning approach that could identify SCI using MRI data. This work concentrates on modeling a novel 2D-convolutional neural networks (2D-CNN) approach for predicting SCI. For holdouts, training, and validation, various datasets of patients were created. Two experts assigned labels to the images. The holdout dataset was used to evaluate the performance of our deep convolutional neural network (DCNN) over the image data from the available dataset. The dataset is acquired from the online resource for training and validation purpose. With the available dataset, the anticipated model attains 94% AUC, 0.1 p-value, and 92.2% accuracy. The anticipated model might make cervical spine MRI scan interpretation more accurate and reliable.

A computational clinical decision-supporting system to suggest effective anti-epileptic drugs for pediatric epilepsy patients based on deep learning models using patient’s medical history

BackgroundEpilepsy, a chronic brain disorder characterized by abnormal brain activity that causes seizures and other symptoms, is typically treated using anti-epileptic drugs (AEDs) as the first-line therapy. However, due to the variations in their modes of action, identification of effective AEDs often relies on ad hoc trials, which is particularly challenging for pediatric patients. Thus, there is significant value in computational methods capable of assisting in the selection of AEDs, aiming to minimize unnecessary medication and improve treatment efficacy.ResultsIn this study, we collected 7,507 medical records from 1,000 pediatric epilepsy patients and developed a computational clinical decision-supporting system for AED selection. This system leverages three multi-channel convolutional neural network (CNN) models tailored to three specific AEDs (vigabatrin, prednisolone, and clobazam). Each CNN model predicts whether a respective AED is effective on a given patient or not. The CNN models showed AUROCs of 0.90, 0.80, and 0.92 in 10-fold cross-validation, respectively. Evaluation on a hold-out test dataset further revealed positive predictive values (PPVs) of 0.92, 0.97, and 0.91 for the three respective CNN models, representing that suggested AEDs by our models would be effective in controlling epilepsy with a high accuracy and thereby reducing unnecessary medications for pediatric patients.ConclusionOur CNN models in the system demonstrated high PPVs for the three AEDs, which signifies the potential of our approach to support the clinical decision-making by assisting doctors in recommending effective AEDs within the three AEDs for patients based on their medical history. This would result in a reduction in the number of unnecessary ad hoc attempts to find an effective AED for pediatric epilepsy patients.