Sukhan Lee

Part annotation of 3D objects is important for part-level analysis, but local geometric evidence may not determine a unique semantic label. We propose an ontology-based annotation refinement framework that represents multiple semantic interpretations and filters candidates using structural and spatial constraints. For each point, a semantic interpretation set is constructed from labels observed in its local neighborhood, and ontology constraints yield a feasible subset without introducing new labels or forcing a single interpretation. We consider boundary ambiguity, part-definition ambiguity, and label-set incompleteness and evaluate the framework on the Airplane, Knife, and Skateboard categories of ShapeNet Part. Since feasible sets are subsets of semantic sets by construction, reductions in Ambiguity Size (AS) and Multi-label Ratio (MLR) are inherent properties of filtering, not independent evidence of improved annotation accuracy. The macro-average Ontology Filtering Rate (OFR) is 0.0947. Point-level analysis distinguishes unchanged candidate sets, partial filtering, and empty feasible sets, characterizing how the adopted ontology constraints affect candidate feasibility. The framework supports ambiguity-aware assessment of existing part annotations while preserving feasible alternatives, without claiming to resolve semantic ambiguity or recover ground-truth labels.

Point cloud distance metrics play an essential role in assessing the quality of point clouds and computing losses in various point cloud processing tasks. The many metrics proposed to date have their own strengths and weaknesses, particularly in their sensitivity to structural or density imbalance and to outlier or point dispersion error traits. In this study, we presented an in-depth analysis of point cloud metrics, examined their sensitivities to density imbalance and point dispersion errors, and demonstrated how these sensitivities affect their performance in point cloud quality assessment and reconstruction. More significantly, we proposed a new metric, the Nested Chamfer Distance (NCD-n), that allows adjusting its sensitivity to density imbalance and point dispersion error traits through the nested CD cycle, n. NCD-n achieves this adjustment by slicing out the one-to-one correspondence point pairs with bilaterally minimum distances, approximating Earth Mover’s Distance (EMD) and allowing the fusion of EMD and CD’s sensitivity properties. In addition, we analyzed the effect of shape-induced point density differences on distance metrics and proposed accounting for shape compensation in these metrics. Through experiments, we verified differences in metric sensitivities to error traits, the effectiveness of the proposed NCD-n in adjusting its sensitivity to density imbalance and point dispersion error traits, and the effects of sensitivity adjustment and shape-compensation on the performance of point cloud quality assessment and reconstruction tasks.

Smartphone-based digital phenotyping has emerged as a promising approach for monitoring mental health using passive behavioral data. Prior studies have linked smartphone-derived features to depression and anxiety severity; however, knowledge regarding whether short-term changes in symptoms can be captured using passive smartphone data in general population samples remains limited, as does the understanding of how such findings should be interpreted vis-à-vis behavioral patterns and demographic variability. This study aimed to model short-term changes in depression and anxiety severity using passive smartphone data, examine model performance across demographic subgroups, and identify behavioral patterns associated with symptom changes. We collected 2 weeks of smartphone usage data from 95 adults in the general population and assessed depressive and anxiety symptoms using the clinician-rated Hamilton Depression Rating Scale and Hamilton Anxiety Rating Scale, respectively. Behavioral features-including physical activity, app use, and screen usage metrics-were extracted and compressed using an autoencoder and principal component analysis. The resulting features-along with age, sex, and baseline Hamilton scores-were used to train random forest classifiers predicting symptom score changes (increase, decrease, or unchanged). Additionally, we examined whether model performance differed across demographic subgroups and whether models excluding baseline scores retained predictive performance, as baseline severity was expected to be a strong predictor. To add explanatory value beyond prediction, behavioral subtypes associated with symptom changes were identified by applying unsupervised clustering. The model exhibited moderate performance in predicting changes in the Hamilton Depression Rating Scale (mean accuracy=0.70, mean area under the receiver operating characteristic curve=0.74) and Hamilton Anxiety Rating Scale (mean accuracy=0.65, mean area under the receiver operating characteristic curve=0.69) scores. Performance varied according to demographics, with reduced accuracy among younger adults and females, although these differences were not significant in permutation tests. Excluding baseline Hamilton scores diminished performance substantially, suggesting that baseline symptom severity accounted for a substantial proportion of the predictive performance. Clustering revealed 4 distinct behavioral subtypes according to smartphone usage patterns. A cluster characterized by structured, daytime-focused smartphone use and lower temporal entropy demonstrated greater improvement in depressive symptoms, whereas clusters with lower and irregular usage patterns exhibited minimal improvement or worsening. Passive smartphone-derived behavioral data demonstrated moderate ability to model short-term symptom changes in this predominantly nonclinical sample. However, a substantial proportion of the predictive performance was attributable to baseline symptom severity, underscoring that passive smartphone data may provide modest supplementary information rather than robust stand-alone predictive value. Nevertheless, clustering analyses indicated that passive data may still assist in identifying behaviorally distinct subtypes associated with different depressive symptom trajectories. These findings reflect a practical contribution to digital phenotyping research by elucidating both the potential and constraints of passive smartphone data for short-term symptom monitoring in small general population samples.

BACKGROUND Smartphone-based digital phenotyping has emerged as a promising approach for monitoring mental health through passive behavioral data. Prior studies have linked smartphone-derived features to depression and anxiety severity; however, several limitations leave such data’s practical utility uncertain. Most investigations have been cross-sectional and depended solely on self-report measures of psychiatric symptoms, and knowledge is limited regarding the influence of demographic differences and how predictive findings should be interpreted. OBJECTIVE This study aimed to model short-term changes in depression and anxiety severity using passive smartphone data, examine model performance across demographic subgroups, and identify behavioral patterns associated with symptom changes. METHODS We collected two weeks of smartphone usage data from 95 adults in the general population and assessed depressive and anxiety symptoms using the clinician-rated Hamilton Depression Rating Scale (HAM-D) and Hamilton Anxiety Rating Scale (HAM-A), respectively. Behavioral features, including physical activity, app use, and screen usage metrics, were extracted and encoded using a deep learning-based autoencoder. Latent representations—reduced using principal component analysis (PCA)—were utilized to train a random forest classifier to predict changes in HAM scores (increase, decrease, or unchanged). Model performance was evaluated using repeated 5-fold cross-validation, while subgroup analyses were conducted by sex and age. Additionally, unsupervised clustering (t-distributed stochastic neighbor embedding followed by K-means) was applied to latent features to further identify explanatory behavioral subtypes. Subsequently, we compared clusters in terms of changes in HAM-A/HAM-D scores and behavioral characteristics, including total usage levels, temporal usage distribution, and behavioral regularity (entropy). RESULTS The model exhibited moderate performance in predicting changes in HAM-D (mean accuracy = 0.70, mean area under the receiver operating characteristic curve [AUC] = 0.74) and HAM-A (mean accuracy = 0.65, mean AUC = 0.69) scores. Performance varied by demographics, with reduced accuracy among younger adults and females. Clustering revealed four distinct behavioral trajectories according to smartphone usage patterns. One cluster, characterized by structured and daytime-concentrated usage with lower temporal entropy, demonstrated significantly greater improvement in depressive symptoms than the others. By contrast, clusters with lower and irregular usage patterns were associated with either minimal improvement or worsening of symptoms. CONCLUSIONS Smartphone-derived behavioral data modestly predict short-term changes in mental health symptoms and uncover distinct usage patterns linked to symptom trajectories. This study’s findings support digital phenotyping’s potential in identifying behaviorally grounded mental health subtypes and informing personalized, data-driven interventions.

BACKGROUND The increasing prevalence of mental health problems among populations prompts the necessity of a mental health self-monitoring platform for the general public as an intermediary to professional treatment, which enables self-diagnosis of mental disorders based on natural lifelog data from daily living environments. OBJECTIVE In this study, we proposed a novel approach to self-predicting depression states with their severity levels based on naturally collected emotional narrative text, referred to here as “sentiment description.” METHODS Unlike conventional text-based approaches, which are mainly lexicon-driven, we extracted holistic mental health clues from the contextual connotations embedded in sentiment descriptions. By integrating them with lexicon-driven features, we aimed to enhance the reliability and explainability of depression diagnosis, supported by rich evidence from various angles. In particular, the holistic mental health clues extracted are designed to have a direct link to the standard mental health indices for diagnosing depression, the level of depression-indicative sentiments associated with chosen aspects, and the personal psychological tendencies affecting the sentiment description-based self-diagnosis network (SD2Net). In SD2Net, 1) holistic clues are extracted based on a large linguistic foundation model, GPT, with optimally designed prompts, 2) the holistic clues and raw texts are transformed into the connotation-driven and lexicon-driven feature embeddings are fused into multi-level depression state prediction, together two learned embeddings are fused into multi-level depression state prediction, together with the regression of their severity-levels, based on a hierarchical decision-making framework. For the experiment, we used both a benchmark dataset and a dataset custom-collected from real subjects on a long-term basis. Notably, we annotated the custom-collected data by fusing multiple self-questionnaires and psychiatrists’ inputs to ensure the trustworthiness of ground truth labels. RESULTS The experiments demonstrate the state-of-the-art performance of the proposed approach for holistic clue-reinforced multi-level depression state prediction, achieving an accuracy of approximately 80% using both a custom-collected dataset and the DepSign benchmark dataset. The ablation study confirms the effectiveness of the proposed holistic clue model, which yields an average improvement of over 10% in accuracy compared to models without it. CONCLUSIONS The proposed approach, SD2Net, shows the effectiveness of incorporating holistic mental health clues from sentiment descriptions for multi-level depression state prediction, highlighting its potential as a reliable self-monitoring tool for the general public.

Point distance metrics such as Chamfer Distance (CD) and Earth Mover’s Distance (EMD) play a key role in evaluating the quality of reconstructed point clouds and computing training loss in point cloud reconstruction. However, the metrics show their own characteristics in evaluation, stemming from differences in their sensitivity to point cloud error traits, which, in turn, causes differences in their performance of point cloud evaluation and reconstruction. In this study, we analyzed in detail the sensitivity of distance metrics to such error traits as density imbalance and point dispersion. We showed how the sensitivity difference in metrics affects their performance in point cloud evaluation and reconstruction. Recognizing that the sensitivity of a metric is correlated with the level of bijective one-to-one point correspondence embedded in the metric, we proposed a new metric, the nestedChamfer distance (NCD-n), that allows to adjust its sensitivity to error traits through the nested CD cycle, n, by which the level of bijective one-to-one point correspondence is controlled. In addition, we analyzed the effect of shape-induced point density difference on distance metrics and proposed taking shape compensation into account in metrics. Based on extensive experiments, we verified the difference of metrics in their sensitivities to error traits, the capability of the proposed NCD-n for adjusting its sensitivity effectively, and the effect of sensitivity adjustment and shape-compensation on the performance of point cloud evaluation and reconstruction.

Noncontact prediction of peripheral oxygen saturation (SpO2) is necessary for monitoring vital signs of patients afflicted by infectious disease or sensitive to skin irritation. Recently, many approaches to imaging photoplethysmography (iPPG)-based noncontact SpO2 prediction have been proposed based on ratio-of-ratio (RoR) representing spectral ratio of diffused absorbances and deep learning (DL). Despite notable progress, issues concerning the accuracy and robustness under real-world variations and artifacts remain to be solved toward clinical viability. In this study, we proposed an approach to highly accurate and robust SpO2 prediction under significant disturbances based on multispectral iPPG signals captured from human facial regions of interest (ROIs). First, a method for accurately estimating actual diffused spectral absorbance of ROIs is proposed. To this end, the shading bias representing the discrepancy between the measured and actual diffused absorbances is calculated by projecting the measured diffused absorbance onto the actual diffused absorbance manifold preformulated based on Monte-Carlo modeling of light transport in multilayered tissues (MCML). Furthermore, the spectral band pairs optimal for estimating shading bias and actual diffused absorbances are selected, resulting in highly accurate RoRs. Second, the SpO2s derived from RoRs are spatiotemporally refined and regressed based on an LSTM AE network further to upgrade the accuracy and robustness of SpO2 prediction. Lastly, individual SpO2s predicted with multiple spectral band pairs and at multiple ROIs are fused to ensure a resilience against possible disruptions due to motion and appearance artifacts. Experiments were conducted with 12 subjects whose SpO2 levels were adjusted between 0.79-0.99 by breath-holding cycles. The cross-validated performance indicated about 2% of Mean Absolute Error (MAE) accuracy in case SpO2s were derived solely from the estimated RoRs while the accuracy of 0.46% of MAE was reached with the LSTM AE-based spatiotemporal refinement and regression.

Background/Objectives: Neuroimaging techniques have been increasingly utilized to explore neuroplasticity induced by various training regimens. Magnetic resonance imaging (MRI) enables to study these changes non-invasively. While visual and motor training have been widely studied, less is known about how auditory training affects brain activity. Our objective was to investigate the effects of sound localization training on brain activity and identify brain regions exhibiting significant changes in activation pre- and post-training to understand how sound localization training induces plasticity in the brain. Method: Six blindfolded participants each underwent 30-minute sound localization training sessions twice a week for three weeks. All participants completed functional MRI (fMRI) testing before and after the training. Results: fMRI scans revealed that sound localization training led to increased activation in several cortical areas, including the superior frontal gyrus, superior temporal gyrus, middle temporal gyrus, parietal lobule, precentral gyrus, and postcentral gyrus. These regions are associated with cognitive processes such as auditory processing, spatial working memory, planning, decision-making, error detection, and motor control. Conversely, a decrease in activation was observed in the left middle temporal gyrus, a region linked to language comprehension and semantic memory. Conclusions: These findings suggest that sound localization training enhances neural activity in areas involved in higher-order cognitive functions, spatial attention, and motor execution, while potentially reducing reliance on regions involved in basic sensory processing. This study provides evidence of training-induced neuroplasticity, highlighting the brain’s capacity to adapt through targeted auditory training intervention.

We were interested in developing a methodology for diagnosing the depression status of a focused population group, such as the Korean university student group, with higher accuracy. To this end, we proposed a method of fusing the data collected from multiple depression self-questionnaires aided by a psychiatrist’s diagnosis. In particular, we found that the standard diagnostic cut-offs and factor analysis prepared for a general population by depression self-questionnaires are inadequate for a focused population with its unique cultural background. In this study, a novel approach to optimizing diagnostic cut-offs and generalizing factor analysis for the Korean university student group is presented in the fusion space of multiple self-questionnaires. We collected the data from 30 randomly selected Korean university students, over 21 weeks, with the psychiatric evaluation as a reference, then established the optimal cut-off regions in the fused CESD − PHQ9 score space based on the statistical correlation between CES − D and PHQ − 9 and the reference diagnostics. We also re-extracted the factors in the fused CESD − PHQ9 space to expose the key factors that are behind the depression characteristics of the group. We verified the existence of a clear correlation between CES − D and PHQ − 9 scores. However, the standard cut-offs of CES − D and PHQ − 9 are found inconsistent with the correlation. The new cut-off regions we obtained in the fused CESD − PHQ9 score space are consistent with the correlation and optimal for the psychiatrist’s diagnosis with the sensitivity and specificity of 80.95% and 89.74%, respectively. Also, we identified that “socio-psychological” and “interpersonal relationship” factors are the major factors for the depression characteristics of the group. Although the new cut-off regions we presented were based on the incorporation of clinical diagnosis into the fused CESD − PHQ9 score space, further verification with a larger scale of clinical data is helpful. We identified optimal cut-off regions and generalized factor analysis in the fusion space, which can provide more reliable and trustworthy diagnoses. These can serve as a self-diagnostic tool for reliably identifying the depression characteristics of a focused population as well as effectively linking individuals and psychiatrists as an intermediary.