""" Formulário de extração de dados — revisão LLMs médicas / Edge AI. Tipos: O=texto, M=texto longo, B=booleano, I=inteiro, S=seleção múltipla. """ from django.db import models from reviews.models import DataExtractionField, DataExtractionLookup # (descrição, tipo, opções ou None) EXTRACTION_FORM: list[tuple[str, str, list[str] | None]] = [ ("Descrição", "M", None), ("Title", "O", None), ("Authors", "O", None), ("Year", "I", None), ("Source", "O", None), ("Peer Reviewed", "B", None), ("Texto completo disponível", "B", None), ( "Application Domain", "S", [ "Biosignal Analysis", "CPS/IoT", "Clinical Decision Support Biomedical", "General Edge AI", "General Healthcare", "Healthcare", "IoT in Healthcare", "Medical Imaging", "Medical Natural Language Processing", "Other", "Smart Hospital", "Smart Office", "Telemedicine", ], ), ("Medical Context", "B", None), ("Dataset/Senário", "O", None), ( "Model Type", "S", [ "Based on CNN", "Decoder-only LLM Model", "Encoder-Decoder Transformer", "Hybrid Architecture", "Mixture of Experts (MoE)", "Multimodal LLM Model", "NNN/LSTM", "Other", "Quantum Neural Network", "Quantum-Inspired Neural Network", "State-Space Model (SSM) / Mamba", "Transformer", "Vision Transformer (ViT)", "Vision-Language Model (VLM)", "Wavelet Neural Network", ], ), ("LLM-based", "B", None), ("Transformer-based", "B", None), ("Wavelet-based Approach", "B", None), ("Quantum-Inspired Approach", "B", None), ("Edge AI Context", "B", None), ("Offline Execution", "B", None), ( "Deployment Architecture", "S", [ "Centralized Cloud", "Cloud-edge Hybrid", "Distributed Edge", "Edge-only", "Federated", "Not Reported", "On-device", ], ), ( "Edge Device Type", "S", [ "Cloud-assisted Edge", "Edge Server", "Embedded Device", "IoT Gateway", "Jetson Nano", "Jetson Orin", "Jetson Xavier", "Laptop-class Edge", "Not Reported", "Other", "Raspberry Pi", "Smartphone", "Wearable", ], ), ("Hardware Platform", "O", None), ( "GPU/CPU/NPU Usage", "S", [ "CPU", "Edge Accelerator", "FPGA", "GPU", "NPU", "Not Reported", "Other", "TPU", ], ), ( "Compression Technique", "S", [ "Entropy Encoding", "Feature Fusion / PCA", "Knowledge Compression", "Low-Rank", "Model Compression", "Model Merging", "None", "PEFT/LoRA", "Prompt Optimization", "Quantization", "Sparse Representation", "Weight Sharing", "Ternary / Sub-byte Quantization", ], ), ( "Quantization Technique", "S", [ "Dynamic Quantization", "FP16", "INT4", "INT8", "Mixed Precision", "None", "Post-training Quantization", "Quantization-aware Training", "Sub-byte Quantization (W2, W3)", "Ternary / Sub-byte Quantization", ], ), ( "Pruning Technique", "S", [ "Asymmetric Pruning", "Layer Pruning", "Magnitude-based", "None", "Structured Pruning", "Token Pruning", "Unstructured Pruning", ], ), ( "Distillation Technique", "S", [ "None", "Progressive Distillation", "Self-Distillation", "Teacher-Student", ], ), ( "Energy Efficiency Metric", "S", [ "Battery Consumption", "CPU Utilization", "Compression Rate", "Energy Consumption", "FLOPs", "GPU Utilization", "Latency", "Memory Footprint", "None", "Power Usage", "RAM Usage", "Thermal Efficiency", "Throughput", "Token Efficiency", ], ), ("Latency Reduction", "O", None), ("Memory Reduction", "O", None), ("Energy Consumption (Watts)", "O", None), ("Model Size (Parameters/Billion)", "O", None), ("Original Model Name", "O", None), ( "Evaluation Type", "S", [ "Ablation Study", "Benchmark", "Case Study", "Comparative Study", "Experimental", "Prototype", "Review", "Simulation", ], ), ("Main Contribution", "M", None), ("Identified Limitations", "M", None), ("Research Gap", "M", None), ( "Efficiency Strategy", "S", [ "Architectural Optimization / Lightweight Design", "Collaborative Inference", "Compression", "Distillation", "Edge Computing", "Low-rank", "Model Merging", "Model Sharding", "None", "Parameter-Efficient Fine-Tuning (PEFT/LoRA)", "Prompt Optimization", "Pruning", "Quantization", "RAG Optimization", "Topic Modeling", ], ), ] FIELD_TYPE_LABELS = { "O": "Texto", "M": "Texto longo", "B": "Sim/Não", "I": "Inteiro", "F": "Decimal", "D": "Data", "S": "Seleção múltipla", } def parse_lookup_options_text(raw: str) -> list[str]: """Uma opção por linha ou separadas por vírgula.""" values: list[str] = [] for line in (raw or "").replace(",", "\n").splitlines(): value = line.strip() if value and value not in values: values.append(value) return values def create_extraction_field( review, description: str, field_type: str = "O", *, lookup_options: list[str] | None = None, ) -> tuple[DataExtractionField | None, str | None]: """ Cria um campo no formulário da revisão. Retorna (campo, mensagem_de_erro). """ desc = (description or "").strip() if not desc: return None, "Informe o nome do campo." ftype = (field_type or "O").strip().upper() if ftype not in DataExtractionField.FIELD_TYPES: return None, "Tipo de campo inválido." existing = DataExtractionField.objects.filter( review_id=review.id, description=desc ).first() if existing: return None, f'Já existe um campo chamado "{desc}".' max_order = ( DataExtractionField.objects.filter(review_id=review.id) .aggregate(max_order=models.Max("order")) .get("max_order") ) order = (max_order if max_order is not None else -1) + 1 field = DataExtractionField( review_id=review.id, description=desc, field_type=ftype, order=order, ) field.save() if ftype == "S" and lookup_options: for value in lookup_options: DataExtractionLookup.objects.create(field_id=field.id, value=value) return field, None def add_extraction_lookup( review, field_id: int, value: str, ) -> tuple[DataExtractionLookup | None, str | None]: """Adiciona uma opção a um campo de seleção múltipla (tipo S).""" field = DataExtractionField.objects.filter( id=field_id, review_id=review.id ).first() if not field: return None, "Campo não encontrado." if field.field_type != "S": return None, "Este campo não é seleção múltipla." val = (value or "").strip() if not val: return None, "Informe o texto da opção." if len(val) > 255: return None, "A opção deve ter no máximo 255 caracteres." existing = DataExtractionLookup.objects.filter(field_id=field.id, value=val).first() if existing: return None, f'Opção "{val}" já existe em "{field.description}".' lookup = DataExtractionLookup(field_id=field.id, value=val) lookup.save() return lookup, None def seed_extraction_form(review, *, replace: bool = True) -> int: """Aplica o formulário padrão da revisão. Retorna quantidade de campos.""" if replace: for field in list(review.extraction_fields.all()): field.delete() for order, (description, field_type, options) in enumerate(EXTRACTION_FORM): field = DataExtractionField( review_id=review.id, description=description, field_type=field_type, order=order, ) field.save() if field_type == "S" and options: for value in options: DataExtractionLookup.objects.create(field_id=field.id, value=value) return len(EXTRACTION_FORM) def ensure_extraction_fields(review) -> int: """Adiciona campos do protocolo que ainda não existem (ex.: Descrição em revisões antigas).""" existing = {f.description for f in review.extraction_fields.all()} added = 0 for order, (description, field_type, options) in enumerate(EXTRACTION_FORM): if description in existing: continue for f in review.extraction_fields.all(): if f.order >= order: f.order += 1 field = DataExtractionField( review_id=review.id, description=description, field_type=field_type, order=order, ) field.save() if field_type == "S" and options: for value in options: DataExtractionLookup.objects.create(field_id=field.id, value=value) existing.add(description) added += 1 if added: for f in DataExtractionField.objects.filter(review_id=review.id).order_by("order"): if f.order is None: f.order = 0 f.save(update_fields=["order"]) return added