JMIR Medical Informatics最新文献

Background: General anesthesia comprises 3 essential components-hypnosis, analgesia, and immobility. Among these, maintaining an appropriate hypnotic state, or anesthetic depth, is crucial for patient safety. Excessively deep anesthesia may lead to hemodynamic instability and postoperative cognitive dysfunction, whereas inadequate anesthesia increases the risk of intraoperative awareness. Electroencephalography (EEG)-based monitoring has therefore become a cornerstone for evaluating anesthetic depth. However, processed electroencephalography (pEEG) indices remain vulnerable to various sources of interference, including electromyographic activity, interindividual variability, and anesthetic drug effects, which can yield inaccurate numerical outputs.

Objective: With recent advances in machine learning, particularly unsupervised learning, data-driven methods that classify signals according to inherent patterns offer new possibilities for anesthetic depth analysis. This study aimed to establish a methodology for automatically identifying anesthesia depth using an unsupervised, machine learning-based clustering approach applied to pEEG data.

Methods: Standard frontal EEG data from participants undergoing elective lumbar spine surgery were retrospectively analyzed, yielding more than 16,000 data points. The signals were filtered with a fourth-order Butterworth bandpass filter and transformed using the fast Fourier transform to estimate power spectral density. Normalized band power ratios for delta, high-theta, alpha, and beta frequencies were extracted as input features. Fuzzy C-Means (FCM) clustering (c=3, m=2) was applied to categorize anesthetic depth into slight, proper, and deep clusters.

Results: FCM clustering successfully identified 3 physiologically interpretable clusters consistent with EEG dynamics during progressive anesthesia. As anesthesia deepened, frontal alpha oscillations became more prominent within a delta-dominant background, while beta activity decreased with loss of consciousness. The fuzzy membership values quantified transitional states and captured the continuum of anesthetic depth. Visualization confirmed strong correspondence among cluster transitions, Patient State Index trends, and spectral density patterns.

Conclusions: This study demonstrates the feasibility of using unsupervised machine learning to enhance anesthetic depth assessment. By applying FCM clustering to pEEG data, this approach improves the understanding of anesthesia depth and integrates effectively with existing monitoring modalities. The proposed FCM-based method complements current EEG indices and may assist anesthesia practitioners and even nonanesthesia professionals in assessing anesthetic depth to enhance patient safety.

Unlabelled: Artificial intelligence (AI) scribes, software that can convert speech into concise clinical documents, have achieved remarkable clinical adoption at a pace rarely seen for digital technologies in health care. The reasons for this are understandable: the technology works well enough, it addresses a genuine pain point for clinicians, and it has largely sidestepped regulatory requirements. In many ways, clinical adoption of AI scribes has also occurred well ahead of robust evidence of their safety and efficacy. The papers in this theme issue demonstrate real progress in the technology and evidence of its benefit: documentation times are reported to decrease when using scribes, clinicians report feeling less burdened, and the notes produced are often of reasonable quality. Yet as we survey the emerging evidence base, there remains one outstanding and urgent unanswered question: Are AI scribes safe? We need to know the clinical outcomes achievable when scribes are used compared to other forms of note taking.

Background: Data linkage in pharmacoepidemiological research is commonly employed to ascertain exposures and outcomes or to obtain additional information on confounding variables. However, to protect patient confidentiality, unique patient identifiers are not provided, which makes data linkage across multiple sources challenging. The Saudi Real-World Evidence Network (SRWEN) aggregates electronic health records from various hospitals, which may require robust linkage techniques.

Objective: We aimed to evaluate and compare the performance of deterministic, probabilistic, and machine learning (ML) approaches for linking deidentified data of patients with multiple sclerosis (MS) from the SRWEN and Ministry of National Guard Health Affairs electronic health record systems.

Methods: A simulation-based validation framework was applied before linking real-world data sources. Deterministic linkage was based on predefined rules, whereas probabilistic linkage was based on a similarity score-based matching. For ML, both similarity score-based and classification approaches were applied using neural networks, logistic regression, and random forest models. The performance of each approach was assessed using confusion matrices, focusing on sensitivity, positive predictive value, F1 score, and computational efficiency.

Results: The study included linked data of 2247 patients with MS from 2016 to 2023. The deterministic approach resulted in an average F1 score of 97.2% in the simulation and demonstrated varying match rates in real-world linkage: 1046/2247 (46.6%) to 1946/2247 (86.6%). This linkage was computationally efficient, with run times of <1 second per rule. The probabilistic approach provided an average F1 score of 93.9% in the simulation, with real-world match rates ranging from 1472/2247 (65.5%) to 2144/2247 (95.4%) and processing times ranging from approximately 0.1 to 5 seconds per rule. ML approaches achieved high performance (F1 score reached 99.8%) but were computationally expensive. Processing times ranged from approximately 13 to 16,936 seconds for the classification-based approaches and from approximately 13 to 7467 seconds for the similarity score-based approaches. Real-world match rates from ML models were highly variable depending on the method used; the similarity score-based approach identified 789/2247 (35.1%) matched pairs, whereas the classification-based approach identified 2014/2247 (89.6%).

Conclusions: Probabilistic linkage offers high linkage capacity by recovering matches missed by deterministic methods and proved to be both flexible and efficient, particularly in real-world scenarios where unique identifiers are lacking. This method achieved a great balance between recall and precision, enabling better integration of various data sources that could be useful in MS research.

Background: The Cox proportional hazards (CPH) model is a common choice for analyzing time-to-treatment interruptions in patients on antiretroviral therapy (ART), valued for its straightforward interpretability and flexibility in handling time-dependent covariates. Machine learning (ML) models have increasingly been adapted for handling temporal data, with added advantages of handling complex, nonlinear relationships and large datasets, and providing clear practical interpretations.

Objective: This study aims to compare the predictive performance of the traditional CPH model and ML models in predicting treatment interruptions among patients on ART, while also providing both global and individual-level explanations to support personalized, data-driven interventions for improving treatment retention.

Methods: Using data from 621,115 patients who started ART between 2017 and 2023, in Kenya, we compared the performance of the CPH with the following ML models-gradient boosting machine, extreme gradient boosting, regularized generalized linear models (Ridge, Lasso, and Elastic-Net), and recursive partitioning-in predicting first and multiple treatment interruptions. Explainable surrogate technique (model-agnostic) was applied to interpret the best performing model's predictions globally, using variable importance and partial dependence profiles, and at individual level, using breakdown additive, Shapley Additive Explanations, and ceteris paribus.

Results: The recursive partitioning model achieved the best performance with a predictive concordance index score of 0.81 for first treatment interruptions and 0.89 for multiple interruptions, outperforming the CPH model, which scored 0.78 and 0.87 for the same scenarios, respectively. Recursive partitioning's performance can be attributed to its ability to model nonlinear relationships and automatically detect complex interactions. The global model-agnostic explanations aligned closely with the interpretations offered by hazard ratios in the CPH model, while offering additional insights into the impact of specific features on the model's predictions. The breakdown additive and Shapley Additive Explanations explainers demonstrated how different variables contribute to the predicted risk at the individual patient level. The ceteris paribus profiles further explored the time-varying model to illustrate how changes in a patient's covariates over time could impact their predicted risk of treatment interruption.

Conclusions: Our results highlight the superior predictive performance of ML models and their ability to provide patient-specific risk predictions and insights that can support targeted interventions to reduce treatment interruptions in ART care.

Background: Multiple instance learning (MIL) is widely used for slide-level classification in digital pathology without requiring expert annotations. However, even partial expert annotations offer valuable supervision; few studies have effectively leveraged this information within MIL frameworks.

Objective: This study aims to develop and evaluate a ranking-aware MIL framework, called rank induction, that effectively incorporates partial expert annotations to improve slide-level classification performance under realistic annotation constraints.

Methods: We developed rank induction, a MIL approach that incorporates expert annotations using a pairwise rank loss inspired by RankNet. The method encourages the model to assign higher attention scores to annotated regions than to unannotated ones, guiding it to focus on diagnostically relevant patches. We evaluated rank induction on 2 public datasets (Camelyon16 and DigestPath2019) and an in-house dataset (Seegene Medical Foundation-stomach; SMF-stomach) and tested its robustness under 3 real-world conditions: low-data regimes, coarse within-slide annotations, and sparse slide-level annotations.

Results: Rank induction outperformed existing methodologies, achieving an area under the receiver operating characteristic curve (AUROC) of 0.839 on Camelyon16, 0.995 on DigestPath2019, and 0.875 on SMF-stomach. It remained robust under low-data conditions, maintaining an AUROC of 0.761 with only 60.2% (130/216) of the training data. When using coarse annotations (with 2240-pixel padding), performance slightly declined to 0.823. Remarkably, annotating just 20% (18/89) of the slides was enough to reach near-saturated performance (AUROC of 0.806, vs 0.839 with full annotations).

Conclusions: Incorporating expert annotations through ranking-based supervision improves MIL-based classification. Rank induction remains robust even with limited, coarse, or sparsely available annotations, demonstrating its practicality in real-world scenarios.

This research letter summarizes early lessons from 4 enterprise implementations of artificial intelligence-enabled customer relationship management platforms in health care and describes governance practices associated with improvements in affordability, adherence, and access at program level.

Background: Rib fractures are present in 10%-15% of thoracic trauma cases but are often missed on chest radiographs, delaying diagnosis and treatment. Artificial intelligence (AI) may improve detection and triage in emergency settings.

Objective: This study aims to evaluate diagnostic accuracy, processing speed, and technical feasibility of an artificial intelligence-assisted rib fracture detection system using prospectively collected data within a real-world, high-volume emergency department workflow.

Methods: We conducted an observational feasibility study with prospective data collection of a faster region-based convolutional neural network-based AI model deployed in the emergency department to analyze 23,251 real-world chest radiographs (22,946 anteroposterior; 305 oblique) from April 1 to July 2, 2023. This study was approved by the Institutional Review Board of MacKay Memorial Hospital (IRB No. 20MMHIS483e). AI operated passively, without influencing clinical decision-making. The reference standard was the final report issued by board-certified radiologists. A subset of discordant cases underwent post hoc computed tomography review for exploratory analysis.

Results: AI achieved 74.5% sensitivity (95% CI 0.708-0.780), 93.3% specificity (95% CI 0.930-0.937), 24.2% positive predictive value, and 99.2% negative predictive value. Median inference time was 10.6 seconds versus 3.3 hours for radiologist reports (paired Wilcoxon signed-rank test W=112 987.5, P<.001). The analysis revealed peak imaging demand between 08:00 and 16:00 and Thursday-Saturday evenings. A 14-day graphics processing unit outage underscored the importance of infrastructure resilience.

Conclusions: The AI system demonstrated strong technical feasibility for real-time rib fracture detection in a high-volume emergency department setting, with rapid inference and stable performance during prospective deployment. Although the system showed high negative predictive value, the observed false-positive and false-negative rates indicate that it should be considered a supportive screening tool rather than a stand-alone diagnostic solution or a replacement for clinical judgment. These findings support further clinician-in-the-loop studies to evaluate clinical feasibility, workflow integration, and impact on diagnostic decision-making. However, interpretation is limited by reliance on radiology reports as the reference standard and the system's passive, non-interventional deployment.