Machine Learning

This multicenter bidirectional cohort study aims to develop and externally validate a machine learning model for predicting the risk of acute exacerbation within 1 year in patients with allergic bronchopulmonary aspergillosis (ABPA) during the stable phase, and further to evaluate the model's practical value in risk stratification and clinical decision-making. All patients diagnosed with ABPA according to the ISHAM 2024 criteria will be assigned to either the acute exacerbation group or the non-exacerbation group based on whether they experience an acute exacerbation within 1 year. Enrolled participants will be randomly divided into a training set and an internal validation set. During the feature selection phase, univariate analysis, collinearity diagnostics, feature importance ranking derived from nine machine learning algorithms, and expert consensus are comprehensively applied, ultimately leading to the development of 12 independent machine learning models. Model performance is assessed using the receiver operating characteristic (ROC) curve and its area under the curve (AUC), sensitivity, specificity, F1-score, calibration curve, and decision curve analysis. In addition, external validation further enhances the credibility of the model. To improve clinical interpretability, the SHAP method is employed to quantify the contribution of each feature, and an interactive nomogram is constructed to facilitate clinical application. All participants will be followed up for 12 months, during which regular clinical and laboratory evaluations will be performed.

Purpose: The primary objective of this study is to assess the diagnostic accuracy of an AI-enabled ultrasound tool for estimating fetal weight Participants: 1,000 pregnant individuals Procedures (methods): This prospective diagnostic accuracy study will enroll 1,000 pregnant individuals within one week of anticipated delivery. At a single visit, each participant will undergo two ultrasound assessments: (1) standardized sweeps for AI analysis (performed by both specialist and nonspecialist users), (2) specialist-performed fetal biometry.

Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by disturbances in communication, poor social skills, and aberrant behaviors. Particularly detrimental are the presence of restricted and repetitive stereotyped behaviors and uncontrollable temper outbursts over trivial changes in the environment, which often cause emotional stress for the children, their families, schools and neighborhood communities. Fundamental to these cognitive and behavioral problems is the disordered cortical connectivity and resultant executive dysfunction that underpin the use of effective strategies to integrate information across contexts. Brain connectivity problems affect the rate at which information travels across the brain. Slow processing speed relates to a reduced capacity of executive function to recall and formulate thoughts and actions automatically, with the result that autistic children with poor processing speed have great difficulty learning or perceiving relationships across multiple experiences. In consequence, these children compensate for the impaired ability to integrate information from the environment by memorizing visual details or individual rules from each situation. This explains why children with autism tend to follow routines in precise detail and show great distress over seemingly trivial changes in the environment. To date, there is no known cure for ASD, and the disorder remains a highly disabling condition. Recently, a non-invasive brain stimulation technique, transcranial direct current Stimulation (tDCS) has shown great promise as a potentially effective and costeffective tool for reducing core symptoms such as anxiety, aggression, impulsivity, and inattention in patients with autism. This technique has been shown to modify behavior by inducing changes in cortical excitability and enhancing connectivity between the targeted brain areas. However, not all ASD patients respond to this intervention the same way and predicting the behavioral impact of tDCS in patients with ASD remains a clinical challenge. This proposed study thus aims to address these challenges by determining whether resting-state EEG and clinical data at baseline can be used to differentiate responders from non-responders to tDCS treatment. Findings from the study will provide new guidance for designing intervention programs for individuals with ASD.

MOSAIC aims to determine whether oro-dental morphological anomalies, particularly palatal morphology, associated with rare bone and cartilage diseases can be precisely characterized using 3D digital models analysed through geometric morphometrics. The study will also evaluate whether these morphological signatures can train an artificial intelligence (AI) algorithm to classify syndromes. A prospective monocentric case-control cohort will be constituted, including 3D intra-oral scans and associated clinical data. The final goal is to improve diagnostic accuracy and reduce diagnostic delay in rare bone disorders.

To develop and externally validate a machine learning model for predicting the 1-year risk of relapse in patients with stable ABPA, and to further evaluate its value in risk stratification and clinical decision-making.

The goal of this observational study is to develop and pre-validate a machine learning algorithm to predict early recovery of mobility in patients undergoing hip or knee joint replacement surgery. The primary research question is: Can a machine learning model accurately classify patients with faster versus slower recovery of autonomous mobility in the first days after joint replacement surgery? Patients who have undergone elective hip or knee arthroplasty and received post-operative physiotherapy will have their clinical and perioperative data collected retrospectively (2020-2023) and prospectively (March 2026-December 2027). The algorithm will be trained on retrospective data and tested prospectively to evaluate its predictive performance for early mobilization and length of hospital stay.

The goal of this trial is to determine the effectiveness of a machine-learning (ML) model predicting a serious cardiac event within the next three months, when compared pre- versus post-deployment, in pediatric cardiac inpatients. The main questions it aims to answer are whether deployment of the ML model: 1. Increases PACT consultation within the next three months among admissions without PACT involvement in the previous 100 days 2. Increases PACT consultation or visit within the next three months among those who experience a serious cardiac event during this period 3. Decreases time to PACT consultation or visit among those seen by PACT during this period 4. Decreases the incidence of death in the intensive care unit (ICU) 5. Increases documentation of goals of care High-risk cardiology patients will be identified by an ML model each morning. If the patient has been seen by the PACT team within the past year, the update will go to the PACT team members. If the patient hasn't been seen by the PACT team, the email will be sent to the cardiology physician in charge of the patient. This physician will decide whether a PACT consultation is necessary based on their clinical judgment. If so, a referral will be made using the usual process. Outcomes of the identified patients will be compared pre- and post-deployment.

Neonatal screening procedures for potentially life-threatening congenital cardiovascular diseases (i.e., duct-dependent systemic or pulmonary circulation), currently implemented at the national level, rely primarily on cardiovascular physical examination performed by a neonatologist. More recently, this approach has been complemented by the assessment of hemoglobin oxygen saturation at both the upper and lower extremities (pre- and post-ductal saturation) in order to improve diagnostic sensitivity, although this practice has not yet been uniformly adopted nationwide. Converging evidence indicates that these screening strategies are affected by significant limitations in both sensitivity (failure to identify affected individuals) and specificity (false-positive findings in healthy subjects). These limitations are associated with substantial overall costs for the healthcare system. Failure to correctly identify affected neonates may result in increased morbidity and mortality, whereas overdiagnosis leads to unnecessary second-level diagnostic investigations and imposes a considerable psychological burden on families, who remain understandably anxious until diagnostic confirmation is achieved. The aim of the present research project (proof-of-concept study) is to develop a digital classifier capable to categorize heart sounds with commercially available digital stethoscopes into a binary classification system distinguishing physiological from pathological sounds. The derivation phase will be followed by a prospective validation phase, in which the classifier will be applied to assess its diagnostic performance. This phase will also evaluate the economic impact of the digital screening approach compared with standard practice. During the derivation phase, neonates with known cardiovascular status, as determined by prior echocardiographic assessment (including both healthy subjects and those with congenital heart disease), will be enrolled. Heart sounds will be recorded in a quiet environment under standard clinical conditions, without sedation. Digital recordings will be stored in WAV format and analyzed to develop a binary classification algorithm capable of distinguishing healthy from pathological cases. Following development, the classifier will be prospectively applied to a validation cohort of neonates undergoing conventional cardiovascular screening (clinical examination and pre- and post-ductal pulse oximetry), followed by classification using the digital tool under investigation. All participants will subsequently undergo confirmatory echocardiography. Diagnostic performance metrics, including sensitivity, specificity, positive and negative predictive values, and likelihood ratios, will be calculated for both the digital and conventional screening modalities. Furthermore, the number of missed pathological cases and the number of unnecessary second-level investigations resulting from false-positive findings will be used to define the economic benefit profile of the proposed screening strategy. Monte Carlo simulation techniques will be employed to extrapolate these findings at the national level, using ISTAT data on birth rates and disease prevalence. It is anticipated that the development of a digital classifier for the binary classification of neonatal heart sounds will be feasible. Moreover, it is expected that this tool will demonstrate superior diagnostic performance compared with current neonatal screening strategies, with beneficial implications not only for the accurate identification of affected and healthy neonates but also for reducing overall healthcare costs associated with missed diagnoses and inappropriate overdiagnosis.

The integration of Artificial Intelligence (AI) in anesthesiology offers the potential to shift patient monitoring from reactive to predictive. Deep learning architectures, specifically Long Short-Term Memory (LSTM) networks, excel at processing complex, time-series data to forecast future clinical states. While standard PK/PD models (such as the state of the art Eleveld model for Propofol and Remifentanil) estimate target-site drug concentrations (Ce), they do not account for real-time, patient-specific dynamic responses. This study aims to deploy an AI framework designed to predict future physiological states.

This study aims to explore the correlation between gastrointestinal blood flow and the incidence of enteral nutrition intolerance (ENI) and its symptoms in critically ill patients, construct and compare predictive models including blood flow parameters, and evaluate their incremental predictive value.

The goal of this trial is to learn if a machine learning (ML) model can help optimize drug therapy in the pediatric population. The main question\[s\] it aims to answer are whether a machine learning model predicting receipt of a targeted medication within the next three months: * Increases the offering of pharmacogenetic testing prior to receipt of a targeted medication * Increases the number of patients with pharmacogenetic results prior to receipt of a targeted medication * Increases the number of patients who have alteration in medication choice or dose based on pharmacogenetic results This trial only focuses on the prediction and provision of participants with a high-risk of receiving a medication with a pharmacogenetic indication in the next three months.

The objective of this study is to determine whether healthcare data and remotely collected patient data can accurately predict asthma attacks in children and young people aged 5-17 years. The main outcome is: when using this new system, is there a reduction in asthma attacks compared with a historic average. The whole population of children and young people with asthma will have routine healthcare data monitored, with a subset of people with high risk asthma asked to participate in a more detail study involving remotely monitored data.

This multicenter, assessor-blinded, two-arm parallel randomized controlled trial (N = 314) will compare the efficacy and safety of a 6-week multidimensional exercise program plus usual pharmacological care (experimental arm) versus usual pharmacological care alone (control arm) in adults ≥ 60 years with chronic non-specific low-back pain (LBP) and imaging evidence of paraspinal muscle degeneration. The primary endpoint is change in Oswestry Disability Index (ODI) at 12 months. Secondary endpoints include pain VAS, JOA score, recurrence rate, and patient satisfaction measured repeatedly to 12 months. Advanced MRI radiomics and machine-learning algorithms will be used to build a "paraspinal muscle imaging-function-prognosis" prediction model and an open-access web tool for risk stratification. The study will generate a standardized, evidence-based non-operative care pathway for chronic LBP driven by paraspinal muscle degeneration

Delirium has a high incidence rate and significantly affects patient prognosis. Diagnosis often relies on manual assessment, which is subject to strong subjectivity, high rates of missed diagnosis, and poor stability. This study employs non-contact identification technology based on machine vision analysis to quantitatively analyze characteristic biological feature data such as micro-expressions. It then investigates the correlation between these features and delirium subtypes. By integrating clinical phenotypic data and using machine learning algorithms, a multi-modal early prediction model for delirium is constructed to meet the clinical need for early warning of delirium subtypes and enhance the efficacy of delirium identification.

The goal of this clinical trial is to use machine learning (ML) to predict functional recovery by integrating muscle-related factors and other relevant parameters for identification of non-responders to conventional rehabilitation. The main questions it aims to answer are: Do deficit clusters lead to poorer functional recovery compared to non-deficit clusters? Does an ML-derived composite score that integrates quadriceps/hamstring strength and size outperform isolated metrics in predicting RTP success? Researchers will compare deficit clusters against non-deficit clusters to determine if deficit clusters lead to poorer functional recovery. Participants will: Return for 5 follow-up timepoints in total for PRO and functional assessments including pre-operation, 1-, 3-, 6- and 12-months post-operation.

1. Background \& Rationale: Accurate assessment of a patient's blood volume (BV) status before surgery is critical for preventing perioperative complications. However, there is currently no clinically feasible, accurate, and non-invasive method for direct BV quantification. We hypothesize that dynamic ultrasound videos of major blood vessels contain rich, sub-visual spatiotemporal information about vascular compliance and filling that can be leveraged to estimate BV. 2. Objective: To develop and validate a deep learning model that integrates multi-modal ultrasound video data to achieve non-invasive, quantitative estimation of preoperative blood volume. 3. Study Design: A prospective, single-center, observational study. 4. Methods: Participants: Adult patients scheduled for surgery. Data Acquisition: Input (Features): Preoperative ultrasound video clips will be recorded in standardized views of four key vessels: the Internal Jugular Vein (IJV), Subclavian Vein (SCV), Inferior Vena Cava (IVC), and Common Carotid Artery (CA). Target (Label): The true Blood Volume (BV) will be calculated for each patient using the acute normovolemic hemodilution (ANH) method. The change in hemoglobin concentration before and after this process is used to calculate the total blood volume with high clinical reliability. Model Development: A hybrid deep learning architecture (e.g., CNN + LSTM/Transformer) will be trained to extract features from the ultrasound videos and learn the complex, non-linear mapping to the BV value derived from ANH. The model will be trained and internally validated using a k-fold cross-validation approach. 5. Expected Outcome \& Significance: We anticipate the development of a novel, end-to-end deep learning model capable of providing a quantitative BV estimate from routine ultrasound scans. This technology has the potential to revolutionize perioperative fluid management by offering a rapid, non-invasive, and accurate tool for objective volume status assessment, ultimately guiding personalized therapy and improving patient outcomes.

Allogeneic hematopoietic stem cell transplantation (HSCT) represents a major therapeutic strategy for malignant hematologic diseases, with the number of procedures steadily increasing in France each year. Conditioning and maintenance regimens carry a risk of both short- and long-term cardiotoxicity, leading to serious cardiovascular events including acute coronary syndrome (ACS), cardiac dysfunction, arrhythmias, pulmonary hypertension, and pericardial effusion. The pathophysiology of cardiotoxicity in HSCT patients remains poorly understood. It is therefore crucial to investigate underlying mechanisms and identify predictive factors of cardiotoxicity in order to provide appropriate cardiological follow-up and management. Current European Society of Cardiology guidelines recommend routine monitoring of HSCT patients with echocardiography and cardiac biomarkers (NT-proBNP, troponin), although these recommendations are based on small-scale studies. The cardiodepressor factor DPP3 has shown promising results in cardio-oncology, with a causal role in anthracycline-induced cardiac dysfunction. Its role in HSCT-related cardiotoxicity requires further evaluation. This multicenter study of HSCT recipients will be a valuable resource, enabling a better understanding of the pathophysiology of cardiotoxicity and prognosis. It will highlight imaging (echocardiography, calcium score, supra-aortic Doppler), electrocardiographic, and biological markers (including DPP3) associated with prognosis.

The primary objective of this observational registry is to develop a comprehensive clinical and imaging score (incorporating echocardiography and cardiac magnetic resonance data) that enhances risk stratification for patients with Takotsubo syndrome. The secondary objectives of this registry are as follows: Investigate the diagnostic value of cardiac magnetic resonance parameters in predicting in-hospital and long-term outcomes in patients with Takotsubo syndrome. Compare the proposed risk stratification score for patients with Takotsubo syndrome with previously existing scores. Investigate the contribution of machine learning models in predicting in-hospital and long-term outcomes compared to standard clinical scores. The design and rationale of this registry are available at 10.1097/RTI.0000000000000709

The goal of this clinical trial is to learn about the ability of non-invasive brain stimulation during sleep to enhance people's deep sleep and its potential benefit on memory in people with mild cognitive impairment via home use sleep therapy device (SleepWISP) as well as learn about biomarkers associated with Alzheimer disease (AD). The clinical trial aims to answer the following main questions: 1. Whether the non-invasive transcranial electrical stimulation (TES) delivered by SleepWISP could provide short-term enhancement of deep sleep in a single night in the target population. 2. Whether TES delivered by SleepWISP could enhance deep sleep over multiple nights in the target population. 3. Whether enhance on deep sleep could improve memory performance in the target population. Participants will be asked to wear non-invasive and painless devices that record their brain activity during sleep along with an actigraphy watch that measures their movement throughout the day. In addition, blood samples or nasal swab assays will be collected from participants multiple times during the study.

The purpose of this study is to highlight the usefulness of artificial intelligence and machine learning to develop computer algorithms that will achieve with great reliability, speed and accuracy the automatic extraction and processing of large volumes of raw and unstructured clinical data from electronic medical files.

This is an observational, low risk tissue based, non-pharmacological, retrospective-prospective study for adults heart and liver transplant patients, related to IRCCS Azienda Ospedaliero-Universitaria di Bologna (IRCCS AOUBO). This clinical study is part of the national multicentric project DARE. The project has the wide overarching aim to develop digital solutions for personalized healthcare.

This clinical trial was designed as a prospective, multicenter, multi-reader multi-case (MRMC), superiority, parallel-controlled study. Participants who met the trial criteria and signed the informed consent form were enrolled. The trial group involved diagnoses of caries on panoramic radiographs using an artificial intelligence-assisted diagnostic system, while the control group involved diagnoses made by dental practitioners specializing in operative dentistry and endodontics with five years of experience, who interpreted oral panoramic radiographs to determine the presence and severity of caries.

The aim of this study is to ensure the safety of patients in a psychiatric clinic and to detect risky behaviors by using machine learning method. Risky behaviors are defined as behaviors that are personally, socially and developmentally undesirable and endanger life and health.Patient safety and maintaining a safe environment are among the primary duties of healthcare professionals. Suicide is the most important evidence-based risk factor, especially among individuals with psychiatric illnesses, and is one of the most important factors that threaten patient safety. At the end of this study, it is aimed to detect risky behaviors of patients before they harm themselves and to enable healthcare professionals to make early intervention for these behaviors, thus supporting a safe treatment environment, with the computer system that has been trained with the machine learning model installed in the clinics.

The goal of this clinical trial is to study whether the use of our blood culture prediction tool is non-inferior to current practice and if it can improve certain outcomes in all adult patients presenting to the emergency department with a clinical indication for a blood culture analysis (according to the treating physician). The primary endpoint is 30-day mortality. Key secondary outcomes are: * hospital admission rates * in-hospital mortality * hospital length-of-stay. In the intervention group, the physician will follow the advice of our blood culture prediction tool. In the comparison group all patients will undergo a blood culture analysis.

Predicting surgical risks are important to patients and clinicians for shared decision making process and management plan. The study team aim to conduct a hybrid type 1 effectiveness implementation study design. A Randomized Controlled Trial where participants undergoing surgery In Singapore General Hospital (SGH) will be allocated in 1:1 ratio to CARES-guided (unblinded to risk level) or to unguided (blinded to risk level) groups. All participants undergoing elective surgeries in SGH will be considered eligible for enrolment into the study. For elective surgeries, the participants will mainly be recruited from Pre-admission Centre. The outcome of this study will help patients and clinicians make better decisions together. Firstly, the deployment of the CARES model in a live clinical environment could potentially reduce postoperative complications and improve the quality of surgical care provision. The findings from this study would allow fine-tuning of CARES as well as further deployment of additional risk models for specific complications other than Mortality and ICU stay. This in turn would translate to better health for the surgical population and improved cost-effectiveness. This is significant as the surgical population is expected to continuously grow due to improved access to care, better technologies and the aging population. Secondly, IMAGINATIVE will be instrumental in improving our understanding of the deployment strategies for AI/ML predictive models in healthcare. Models such as CARES could be the standard of care in the future if proven to improve the health outcomes of patients. As model deployments are costly and can be disruptive to the EMR processes, this study would be the initial spark for future deployment and health services research focusing on improving the value of these model deployments.