Show me evidence of Model Evaluation
Here’s the evidence.
Poultry Disease Classification: VGG16 transfer learning with targeted fine-tuning; 94% validation accuracy across four categories. 94% Validation accuracy · local research. Four poultry disease categories; dataset identity and sample count not supplied. Four-class VGG16 transfer learning with targeted fine-tuning using TensorFlow/Keras. Validation accuracy across four classes; dataset size, split details, class names and external validation are not supplied. Sub-500 ms inference statement has no supplied hardware, percentile, sample size or production SLA definition. Swin Transformer for EEG Seizure Classification: Validated recording/annotation handling, filtering, segmentation, parameter estimation, reconstruction, fuzzy clustering and real-EEG FRP checks. Selected annotated TUSZ recordings. 19 EEG channels; delta/theta/alpha/beta/gamma filtering; non-overlapping 1-second seizure windows; per-segment AMI/FNN and signal-dependent fuzzy threshold; phase-space reconstruction and fuzzy C-means. Four-class segment discrimination remains challenging; accuracy is modest across bands. MYSZ has only one patient and is excluded from the patient-separated pilot. Only five held-out test patients; initial generalization pilot, not a definitive cross-patient performance estimate. Segment and patient results are not directly comparable: class sets, evaluation units and split strategies differ. Controlled CNN-vs-Swin patient-separated comparison is ongoing/planned; it has no supplied completed results. Internal Biomedical Sensors and Systems Lab, University of Memphis research; no public repository, publication URL or stable external artifact link is supplied. AutoCare Agent: Hybrid BM25 + all-MiniLM-L6-v2 dense retrieval evaluated across 100 diagnostic queries. 90% Recall@5 · local research. 100 diagnostic queries; 20 service manuals; approximately 4,500 document chunks. OpenSearch BM25 + all-MiniLM-L6-v2 dense embeddings through Sentence-Transformers. Decision-support POC, not a substitute for a certified technician. Recall describes the 100-query retrieval evaluation, not field diagnostic accuracy. Separate lexical/dense scores and scenario pass rates are not supplied. EEG Signal Processing & Classification: Sleep Spindle Detection and Motor Imagery Classification: Baseline CSP classifiers achieve approximately 61.5% / 61.7% mean session-wise LOSO accuracy. Baseline CSP + LDA and CSP + linear SVM mean LOSO accuracy. ~61.5% Baseline CSP + LDA LOSO accuracy · local research. PhysioNet EEG Motor Movement/Imagery: clean 106-subject baseline, sessions R04, R08 and R12. 8–30 Hz filtering; cue-locked epochs -1.0 to 4.0 s, baseline -1.0 to 0.0 s, classification window 0.5 to 3.5 s. Fit CSP on training folds only. Subject-wise Leave-One-Session-Out across R04, R08 and R12, avoiding random trial splits. Research workflows and benchmarks, not clinical diagnostic systems or production medical devices. EPCTL11 detection-count agreement is a calibration-participant result, not universal detection improvement or validated accuracy, sensitivity or specificity. EPCTL14 generalized well; EPCTL12 showed moderate over-sensitivity; EPCTL13 and EPCTL15 exposed cross-subject failure modes. Manual detections generally had shorter average durations than YASA. No aggregate cross-participant quality metric is supplied. S088, S092 and S100 were excluded because their 128 Hz recordings differed from the expected 160 Hz; the clean baseline contains 106 of 109 audited subjects. LOSO means Leave-One-Session-Out within subjects across R04, R08 and R12, not leave-one-subject-out or clinical performance. Tuning cohort size and uncertainty estimates are not separately supplied. Tuning yielded only about 1–1.2 percentage points of reported mean LOSO improvement; subject/session non-stationarity remained a larger constraint than classifier choice alone. ANPHY-Sleep / PhysioNet methods, metrics and artifact belong to this project, separately from Swin/FRP seizure classification on TUSZ. EEG Signal Processing & Classification: Sleep Spindle Detection and Motor Imagery Classification: Tuned CSP classifiers achieve approximately 62.7% / 62.4% mean session-wise LOSO accuracy. Tuned CSP + LDA and CSP + linear SVM mean LOSO accuracy; no inferred tuning cohort size. ~62.7% Tuned CSP + LDA LOSO accuracy · local research. PhysioNet EEG Motor Movement/Imagery left-vs-right motor imagery; tuning cohort size not separately supplied. 8–30 Hz filtering; cue-locked epochs -1.0 to 4.0 s, baseline -1.0 to 0.0 s, classification window 0.5 to 3.5 s. Fit CSP on training folds only. Subject-wise Leave-One-Session-Out across R04, R08 and R12, avoiding random trial splits. LDA: 6 CSP components, LSQR, shrinkage = 0.3. Linear SVM: 8 CSP components. Confusion matrices, CSP spatial patterns and CSV summaries inspected in a Streamlit experiment dashboard. Research workflows and benchmarks, not clinical diagnostic systems or production medical devices. EPCTL11 detection-count agreement is a calibration-participant result, not universal detection improvement or validated accuracy, sensitivity or specificity. EPCTL14 generalized well; EPCTL12 showed moderate over-sensitivity; EPCTL13 and EPCTL15 exposed cross-subject failure modes. Manual detections generally had shorter average durations than YASA. No aggregate cross-participant quality metric is supplied. S088, S092 and S100 were excluded because their 128 Hz recordings differed from the expected 160 Hz; the clean baseline contains 106 of 109 audited subjects. LOSO means Leave-One-Session-Out within subjects across R04, R08 and R12, not leave-one-subject-out or clinical performance. Tuning cohort size and uncertainty estimates are not separately supplied. Tuning yielded only about 1–1.2 percentage points of reported mean LOSO improvement; subject/session non-stationarity remained a larger constraint than classifier choice alone. ANPHY-Sleep / PhysioNet methods, metrics and artifact belong to this project, separately from Swin/FRP seizure classification on TUSZ.