The lack of explainability in using relevant clinical knowledge hinders the adoption of artificial intelligence-powered analysis of unstructured clinical dialogue. A wealth of relevant, untapped Mental Health (MH) data is available in online communities, providing the opportunity to address the explainability problem with substantial potential impact as a screening tool for both online and offline applications. Inspired by how clinicians rely on their expertise when interacting with patients, we leverage relevant clinical knowledge to classify and explain depression-related data, reducing manual review time and engendering trust. We developed a method to enhance attention in contemporary transformer models and generate explanations for classifications that are understandable by mental health practitioners (MHPs) by incorporating external clinical knowledge. We propose a domain-general architecture called ProcesS knowledgeinfused cross ATtention (PSAT) that incorporates clinical practice guidelines (CPG) when computing attention. We transform a CPG resource focused on depression, such as the Patient Health Questionnaire (e.g. PHQ-9) and related questions, into a machine-readable ontology using SNOMED-CT. With this resource, PSAT enhances the ability of models like GPT-3.5 to generate application-relevant explanations. Evaluation of four expert-curated datasets related to depression demonstrates PSAT's applicationrelevant explanations. PSAT surpasses the performance of twelve baseline models and can provide explanations where other baselines fall short.
11 April 2025 in Proceedings of the AAAI Conference on Artificial Intelligence
Malware authors often employ code obfuscations to make their malware harder to detect. Existing tools for generating obfuscated code often require access to the original source code (e.g., C++ or Java), and adding new obfuscations is a non-trivial, labor-intensive process. In this study, we ask the following question: Can Large Language Models (LLMs) potentially generate a new obfuscated assembly code? If so, this poses a risk to anti-virus engines and potentially increases the flexibility of attackers to create new obfuscation patterns. We answer this in the affirmative by developing the MetamorphASM benchmark comprising MetamorphASM Dataset (MAD) along with three code obfuscation techniques: dead code, register substitution, and control flow change. The MetamorphASM systematically evaluates the ability of LLMs to generate and analyze obfuscated code using MAD, which contains 328,200 obfuscated assembly code samples. We release this dataset and analyze the success rate of various LLMs (e.g., GPT-3.5/4, GPT-4o-mini, Starcoder, CodeGemma, CodeLlama, CodeT5, and LLaMA 3.1) in generating obfuscated assembly code. The evaluation was performed using established information-theoretic metrics and manual human review to ensure correctness and provide the foundation for researchers to study and develop remediations to this risk.
Modeling is essential to better understand the generative mechanisms responsible for experimental observations gathered from complex systems. In this work, we are using such an approach to analyze the electrocardiogram (ECG). We present a systematic framework to decompose ECG signals into sums of overlapping lognormal components. We use reinforcement learning to train a deep neural network to estimate the modeling parameters from an ECG recorded in babies from 1 to 24 months of age. We demonstrate this model-driven approach by showing how the extracted parameters vary with age. From the 751,510 PQRST complexes modeled, 82.7% provided a signal-to-noise ratio that was sufficient for further analysis (>5 dB). After correction for multiple tests, 10 of the 24 modeling parameters exhibited statistical significance below the 0.01 threshold, with absolute Kendall rank correlation coefficients in the [0.27, 0.51] range. These results confirm that this model-driven approach can capture sensitive ECG parameters. Due to its physiological interpretability, this approach can provide a window into latent variables which are important for understanding the heart-beating process and its control by the autonomous nervous system.
Attribution in large language models (LLMs) remains a significant challenge, particularly in ensuring the factual accuracy and reliability of the generated outputs. Current methods for citation or attribution, such as those employed by tools like Perplexity.ai and Bing-Search-integrated LLMs, attempt to ground responses by providing real-time search results and citations. However, so far, these approaches suffer from issues such as hallucinations; biases; surface-level relevance matching; and the complexity of managing vast, unfiltered knowledge sources. While tools like Perplexity.ai dynamically integrate Web-based information and citations, they often rely on inconsistent sources, such as blog posts or unreliable sources, which limits their overall reliability. We explain that these challenges can be mitigated by integrating neurosymbolic AI (NesyAI), which combines the strengths of neural networks, with structured symbolic reasoning. NesyAI offers transparent, interpretable, and dynamic reasoning processes, addressing the limitations of current attribution methods by incorporating structured symbolic knowledge with flexible, neural-based learning. This article explores how NesyAI frameworks can enhance existing attribution models, offering more reliable, interpretable, and adaptable systems for LLMs.
In recent years, there has been a rise in the prevalence of autism spectrum disorder (ASD). The diagnosis of ASD requires behavioral observation and standardized testing completed by highly trained experts. Early intervention for ASD can begin as early as 1–2 years of age, but ASD diagnoses are not typically made until ages 2–5 years, thus delaying the start of intervention. There is an urgent need for non-invasive biomarkers to detect ASD in infancy. While previous research using physiological recordings has focused on brain-based biomarkers of ASD, this study investigated the potential of electrocardiogram (ECG) recordings as an ASD biomarker in 3–6-month-old infants. We recorded the heart activity of infants at typical and elevated familial likelihood for ASD during naturalistic interactions with objects and caregivers. After obtaining the ECG signals, features such as heart rate variability (HRV) and sympathetic and parasympathetic activities were extracted. Then we evaluated the effectiveness of multiple machine learning classifiers for classifying ASD likelihood. Our findings support our hypothesis that infant ECG signals contain important information about ASD familial likelihood. Amongthe various machine learning algorithms tested, KNN performed best according to sensitivity (0.70 ± 0.117), F1-score (0.689 ± 0.124), precision (0.717 ± 0.128), accuracy (0.70 ± 0.117, p-value = 0.02), and ROC (0.686 ± 0.122, p-value = 0.06). These results suggest that ECG signals contain relevant information about the likelihood of an infant developing ASD. Future studies should consider the potential of information contained in ECG, and other indices of autonomic control, for the development of biomarkers of ASD in infancy.
Several decades of research have investigated the neural connections between stroke-induced brain damage and language difficulties. Typically, lesion-symptom mapping (LSM) studies that address this connection have relied on mass univariate statistics, which do not account for multidimensional relationships between variables. Machine learning (ML) techniques, which can capture these intricate connections, offer a promising complement to LSM methods. To test this promise, we benchmarked ML models on structural and functional MRI to predict aphasia severity (N = 238) and naming impairment (N = 191) for a cohort of chronic-stage stroke survivors. We used nested cross-validation to examine performance along three dimensions: (1) parcellation schemes (JHU, AAL, BRO, and AICHA atlases), (2) neuroimaging modalities (resting-state functional connectivity, structural connectivity, mean diffusivity, fractional anisotropy, and lesion location) and (3) ML methods (Random Forest, Support Vector Regression, Decision Tree, K Nearest Neighbors, and Gradient Boosting). The best results were obtained by combining the JHU atlas, lesion location, and the Random Forest model. This combination yielded moderate to high correlations with the two different behavioral scores. Key regions identified included several perisylvian areas and pathways within the language network. This work complements existing LSM methods with new tools for improving the prediction of language outcomes in stroke survivors.