AI CURATION

# AI-Assisted Curation Reference Guide to using AI visual analysis for unit curation, inspired by SpikeAgent's approach. ## Overview AI-assisted curation uses vision-language models to analyze spike sorting visualizations, providing expert-level quality assessments similar to human curators. ### Workflow ``` Traditional: Metrics → Threshold → Labels AI-Enhanced: Metrics → AI Visual Analysis → Confidence Score → Labels ``` ## Claude Code Integration When using this skill within Claude Code, Claude can directly analyze waveform plots without requiring API setup. Simply: 1. Generate a unit report or plot 2. Ask Claude to analyze the visualization 3. Claude will provide expert-level curation decisions Example workflow in Claude Code: ```python # Generate plots for a unit npa.plot_unit_summary(analyzer, unit_id=0, output='unit_0_summary.png') # Then ask Claude: "Please analyze this unit's waveforms and autocorrelogram # to determine if it's a well-isolated single unit, multi-unit activity, or noise" ``` Claude can assess: - Waveform consistency and shape - Refractory period violations from autocorrelograms - Amplitude stability over time - Overall unit isolation quality ## Quick Start ### Generate Unit Report ```python import neuropixels_analysis as npa # Create visual report for a unit report = npa.generate_unit_report(analyzer, unit_id=0, output_dir='reports/') # Report includes: # - Waveforms, templates, autocorrelogram # - Amplitudes over time, ISI histogram # - Quality metrics summary # - Base64 encoded image for API ``` ### AI Visual Analysis ```python from anthropic import Anthropic # Setup API client client = Anthropic() # Analyze single unit result = npa.analyze_unit_visually( analyzer, unit_id=0, api_client=client, model='claude-3-5-sonnet-20241022', task='quality_assessment' ) print(f"Classification: {result['classification']}") print(f"Reasoning: {result['reasoning']}") ``` ### Batch Analysis ```python # Analyze all units results = npa.batch_visual_curation( analyzer, api_client=client, output_dir='ai_curation/', progress_callback=lambda i, n: print(f"Progress: {i}/{n}") ) # Get labels ai_labels = {uid: r['classification'] for uid, r in results.items()} ``` ## Interactive Curation Session For human-in-the-loop curation with AI assistance: ```python # Create session session = npa.CurationSession.create( analyzer, output_dir='curation_session/', sort_by_confidence=True # Show uncertain units first ) # Process units while True: unit = session.current_unit() if unit is None: break print(f"Unit {unit.unit_id}:") print(f" Auto: {unit.auto_classification} (conf: {unit.confidence:.2f})") # Generate report report = npa.generate_unit_report(analyzer, unit.unit_id) # Get AI opinion ai_result = npa.analyze_unit_visually(analyzer, unit.unit_id, api_client=client) session.set_ai_classification(unit.unit_id, ai_result['classification']) # Human decision decision = input("Decision (good/mua/noise/skip): ") if decision != 'skip': session.set_decision(unit.unit_id, decision) session.next_unit() # Export results labels = session.get_final_labels() session.export_decisions('final_curation.csv') ``` ## Analysis Tasks ### Quality Assessment (Default) Analyzes waveform shape, refractory period, amplitude stability. ```python result = npa.analyze_unit_visually(analyzer, uid, task='quality_assessment') # Returns: 'good', 'mua', or 'noise' ``` ### Merge Candidate Detection Determines if two units should be merged. ```python result = npa.analyze_unit_visually(analyzer, uid, task='merge_candidate') # Returns: 'merge' or 'keep_separate' ``` ### Drift Assessment Evaluates motion/drift in the recording. ```python result = npa.analyze_unit_visually(analyzer, uid, task='drift_assessment') # Returns drift magnitude and correction recommendation ``` ## Custom Prompts Create custom analysis prompts: ```python from neuropixels_analysis.ai_curation import create_curation_prompt # Get base prompt prompt = create_curation_prompt( task='quality_assessment', additional_context='Focus on waveform amplitude consistency' ) # Or fully custom custom_prompt = """ Analyze this unit and determine if it represents a fast-spiking interneuron. Look for: 1. Narrow waveform (peak-to-trough 10 and row['isi_violations_ratio'] < 0.001: labels[unit_id] = 'good' continue if row['snr'] AI > auto) labels = session.get_final_labels() # Export detailed results df = session.export_decisions('curation_results.csv') # Summary summary = session.get_summary() print(f"Good: {summary['decisions'].get('good', 0)}") print(f"MUA: {summary['decisions'].get('mua', 0)}") print(f"Noise: {summary['decisions'].get('noise', 0)}") ``` ## Visual Report Components The generated report includes 6 panels: | Panel | Content | What to Look For | |-------|---------|------------------| | Waveforms | Individual spike waveforms | Consistency, shape | | Template | Mean ± std | Clean negative peak, physiological shape | | Autocorrelogram | Spike timing | Gap at 0ms (refractory period) | | Amplitudes | Amplitude over time | Stability, no drift | | ISI Histogram | Inter-spike intervals | Refractory gap 2 and snr 0.001 and isi_violations_ratio < 0.1 """).index.tolist() # Batch process only these results = npa.batch_visual_curation( analyzer, unit_ids=uncertain_units, api_client=client ) ``` ## References - [SpikeAgent](https://github.com/SpikeAgent/SpikeAgent) - AI-powered spike sorting assistant - [Anthropic Vision API](https://docs.anthropic.com/en/docs/vision) - [GPT-4 Vision](https://platform.openai.com/docs/guides/vision)