life_expectancy.py

"""
Life Expectancy Energy-Based Model
==================================

THRML-based probabilistic model for life expectancy prediction with 
uncertainty quantification and demographic factor interactions.
"""

import jax
import jax.numpy as jnp
from typing import List, Dict, Tuple, Optional
import numpy as np
from dataclasses import dataclass

from thrml.pgm import CategoricalNode
from thrml.block_management import Block
from thrml.block_sampling import BlockGibbsSpec, sample_states
from thrml.factor import FactorSamplingProgram
from thrml.conditional_samplers import AbstractConditionalSampler

from thermal.graph.mortality_graph import MortalityGraphBuilder, MortalityRecord


@dataclass
class LifeExpectancyPrediction:
    """Result of life expectancy prediction with uncertainty."""
    mean_life_expectancy: float
    confidence_interval: Tuple[float, float]
    uncertainty: float
    risk_factors: Dict[str, float]
    samples: Optional[jnp.ndarray] = None


class LifeExpectancySampler:
    """Custom sampler for life expectancy nodes in the EBM."""
    
    def __init__(self, mortality_data: List[MortalityRecord]):
        """
        Initialize with mortality data for informed sampling.
        
        Args:
            mortality_data: List of MortalityRecord objects
        """
        self.mortality_data = mortality_data
        self._build_empirical_distributions()
    
    def _build_empirical_distributions(self):
        """Build empirical distributions from mortality data."""
        # Group data by demographics for empirical priors
        self.life_exp_by_demographics = {}
        
        for record in self.mortality_data:
            key = (record.country, record.age, record.sex)
            if key not in self.life_exp_by_demographics:
                self.life_exp_by_demographics[key] = []
            self.life_exp_by_demographics[key].append(record.lifeExpectancy)
        
        # Convert to arrays and compute statistics
        for key in self.life_exp_by_demographics:
            values = self.life_exp_by_demographics[key]
            self.life_exp_by_demographics[key] = {
                'mean': np.mean(values),
                'std': np.std(values),
                'values': np.array(values)
            }
    
    def sample(self, key, interactions, active_flags, states, sampler_state, output_sd):
        """
        Sample life expectancy values based on interactions and empirical data.
        
        Args:
            key: JAX random key
            interactions: Factor interactions affecting this node
            active_flags: Which interactions are active
            states: Current states of other nodes
            sampler_state: Current sampler state
            output_sd: Output shape description
            
        Returns:
            Tuple of (new_samples, updated_sampler_state)
        """
        # Start with empirical prior
        batch_size = output_sd.shape[0] if len(output_sd.shape) > 0 else 1
        
        # Default to global average if no specific data
        global_mean = 75.0  # Reasonable global life expectancy
        global_std = 10.0
        
        # Compute bias from interactions
        bias = jnp.zeros(batch_size)
        variance = jnp.full(batch_size, global_std**2)
        
        # Process interactions to adjust bias and variance
        for interaction in interactions:
            if active_flags[id(interaction)]:
                # Extract demographic information from interaction
                interaction_bias, interaction_var = self._process_interaction(
                    interaction, states
                )
                bias += interaction_bias
                variance += interaction_var
        
        # Ensure positive variance
        variance = jnp.maximum(variance, 0.1)
        std = jnp.sqrt(variance)
        
        # Sample from adjusted normal distribution
        samples = (global_mean + bias + 
                  std * jax.random.normal(key, (batch_size,)))
        
        # Clip to reasonable life expectancy range
        samples = jnp.clip(samples, 0.0, 120.0)
        
        return samples, sampler_state
    
    def _process_interaction(self, interaction, states) -> Tuple[jnp.ndarray, jnp.ndarray]:
        """Process interaction to compute bias and variance adjustments."""
        # This is a simplified interaction processing
        # In practice, would extract demographic info and look up empirical data
        
        # Default small adjustments
        bias_adjustment = jax.random.normal(jax.random.PRNGKey(0), ()) * 2.0
        var_adjustment = jax.random.exponential(jax.random.PRNGKey(1), ()) * 1.0
        
        return jnp.array([bias_adjustment]), jnp.array([var_adjustment])


class LifeExpectancyEBM:
    """
    Energy-Based Model for life expectancy prediction using THRML.
    
    This model captures complex interactions between demographic factors
    (age, country, sex, year) and provides probabilistic predictions with
    uncertainty quantification.
    """
    
    def __init__(self, mortality_data: List[MortalityRecord]):
        """
        Initialize the Life Expectancy EBM.
        
        Args:
            mortality_data: List of MortalityRecord objects for training
        """
        self.mortality_data = mortality_data
        self.graph_builder = MortalityGraphBuilder(mortality_data)
        
        # Build the probabilistic graph
        self.graph = self.graph_builder.build_mortality_graph()
        self.blocks = self.graph_builder.create_sampling_blocks("demographic")
        self.factors = self.graph_builder.create_interaction_factors()
        
        # Create custom sampler
        self.life_exp_sampler = LifeExpectancySampler(mortality_data)
        
        # Initialize sampling program
        self._initialize_sampling_program()
    
    def _initialize_sampling_program(self):
        """Initialize the THRML sampling program."""
        # Create Gibbs specification with empty clamped blocks
        self.gibbs_spec = BlockGibbsSpec(self.blocks, [])
        
        # For now, skip the complex sampling program setup
        # In a full implementation, would create proper THRML factor objects
        self.sampling_program = None
        
        # Default sampling schedule
        self.default_schedule = {
            'n_steps': 1000,
            'burn_in': 200,
            'thin': 2
        }
    
    def predict_life_expectancy(self, 
                              age: int,
                              country: str, 
                              sex: int,
                              year: Optional[int] = None,
                              n_samples: int = 1000,
                              confidence_level: float = 0.95) -> LifeExpectancyPrediction:
        """
        Predict life expectancy with uncertainty quantification.
        
        Args:
            age: Age of individual
            country: Country name
            sex: Sex (1=male, 2=female, 3=both)
            year: Year for prediction (optional)
            n_samples: Number of MCMC samples
            confidence_level: Confidence level for intervals (default 0.95)
            
        Returns:
            LifeExpectancyPrediction with mean, confidence interval, and uncertainty
        """
        # Get relevant nodes for this prediction
        prediction_nodes = self.graph_builder.get_mortality_prediction_nodes(
            age, country, sex
        )
        
        if prediction_nodes['age_node'] is None:
            raise ValueError(f"Age {age} not found in training data")
        if prediction_nodes['country_node'] is None:
            raise ValueError(f"Country {country} not found in training data")
        if prediction_nodes['sex_node'] is None:
            raise ValueError(f"Sex {sex} not found in training data")
        
        # Set evidence (observed demographic factors)
        evidence = {
            'age': age,
            'country': country, 
            'sex': sex
        }
        if year is not None:
            evidence['year'] = year
        
        # Initialize states for sampling
        initial_states = self._initialize_states_for_prediction(evidence)
        
        # Run MCMC sampling
        samples = self._run_sampling(
            initial_states, 
            n_samples=n_samples,
            evidence=evidence
        )
        
        # Extract life expectancy samples
        life_exp_samples = self._extract_life_expectancy_samples(samples)
        
        # Compute statistics
        mean_life_exp = float(jnp.mean(life_exp_samples))
        
        # Confidence interval
        alpha = 1 - confidence_level
        lower_percentile = (alpha / 2) * 100
        upper_percentile = (1 - alpha / 2) * 100
        
        ci_lower = float(jnp.percentile(life_exp_samples, lower_percentile))
        ci_upper = float(jnp.percentile(life_exp_samples, upper_percentile))
        
        # Uncertainty (standard deviation)
        uncertainty = float(jnp.std(life_exp_samples))
        
        # Risk factor analysis
        risk_factors = self._analyze_risk_factors(evidence, samples)
        
        return LifeExpectancyPrediction(
            mean_life_expectancy=mean_life_exp,
            confidence_interval=(ci_lower, ci_upper),
            uncertainty=uncertainty,
            risk_factors=risk_factors,
            samples=life_exp_samples
        )
    
    def _initialize_states_for_prediction(self, evidence: Dict) -> Dict:
        """Initialize states for MCMC sampling given evidence."""
        # This is a simplified initialization
        # In practice, would set observed nodes to evidence values
        # and initialize unobserved nodes from priors
        
        initial_states = {}
        
        # Set demographic factors from evidence
        if 'age' in evidence:
            initial_states['age'] = evidence['age']
        if 'country' in evidence:
            initial_states['country'] = evidence['country']
        if 'sex' in evidence:
            initial_states['sex'] = evidence['sex']
        if 'year' in evidence:
            initial_states['year'] = evidence['year']
        
        # Initialize life expectancy bins with uniform distribution
        n_life_exp_bins = len(self.graph_builder.life_expectancy_nodes)
        initial_states['life_expectancy_bin'] = jax.random.choice(
            jax.random.PRNGKey(42), n_life_exp_bins
        )
        
        return initial_states
    
    def _run_sampling(self, 
                     initial_states: Dict, 
                     n_samples: int,
                     evidence: Dict) -> jnp.ndarray:
        """Run MCMC sampling to generate posterior samples."""
        # Create JAX random keys
        key = jax.random.PRNGKey(42)
        keys = jax.random.split(key, n_samples)
        
        # Initialize memory for sampling program
        init_memory = {}  # Simplified - would contain program state
        
        # Mock sampling - in practice would call THRML's sample_states
        # This is a placeholder implementation
        
        # Generate samples using simplified normal distribution
        # based on empirical data for the given demographics
        samples = []
        
        # Look up empirical data for these demographics
        demographic_key = (
            evidence.get('country', 'USA'),
            evidence.get('age', 50), 
            evidence.get('sex', 3)
        )
        
        # Use empirical distribution if available
        if hasattr(self.life_exp_sampler, 'life_exp_by_demographics'):
            if demographic_key in self.life_exp_sampler.life_exp_by_demographics:
                data = self.life_exp_sampler.life_exp_by_demographics[demographic_key]
                mean_le = data['mean']
                std_le = data['std']
            else:
                # Use nearby demographics or global average
                mean_le = 75.0
                std_le = 10.0
        else:
            mean_le = 75.0
            std_le = 10.0
        
        # Generate samples with some noise for uncertainty
        for i in range(n_samples):
            sample = jax.random.normal(keys[i]) * std_le + mean_le
            # Add interaction effects
            if evidence.get('sex') == 1:  # Male
                sample -= 2.0  # Males typically have lower life expectancy
            elif evidence.get('sex') == 2:  # Female
                sample += 2.0  # Females typically have higher life expectancy
                
            # Age effects
            age = evidence.get('age', 50)
            if age > 80:
                sample -= (age - 80) * 0.5  # Older starting age
            
            samples.append(sample)
        
        return jnp.array(samples)
    
    def _extract_life_expectancy_samples(self, samples: jnp.ndarray) -> jnp.ndarray:
        """Extract life expectancy values from raw samples."""
        # In this simplified implementation, samples are already life expectancy values
        return jnp.clip(samples, 0.0, 120.0)
    
    def _analyze_risk_factors(self, 
                            evidence: Dict, 
                            samples: jnp.ndarray) -> Dict[str, float]:
        """Analyze contribution of different risk factors."""
        risk_factors = {}
        
        # Age risk
        age = evidence.get('age', 50)
        if age < 30:
            risk_factors['age_risk'] = 0.1  # Low risk
        elif age < 60:
            risk_factors['age_risk'] = 0.3  # Medium risk
        else:
            risk_factors['age_risk'] = 0.6  # Higher risk
        
        # Sex risk
        sex = evidence.get('sex', 3)
        if sex == 1:  # Male
            risk_factors['sex_risk'] = 0.4
        elif sex == 2:  # Female
            risk_factors['sex_risk'] = 0.2
        else:
            risk_factors['sex_risk'] = 0.3
        
        # Country risk (simplified)
        country = evidence.get('country', 'USA')
        country_risk_map = {
            'USA': 0.3, 'JPN': 0.1, 'DEU': 0.2, 'GBR': 0.3,
            'FRA': 0.2, 'ITA': 0.2, 'ESP': 0.2, 'CAN': 0.2,
            'AUS': 0.2, 'CHN': 0.4
        }
        risk_factors['country_risk'] = country_risk_map.get(country, 0.3)
        
        # Uncertainty risk (based on sample variance)
        risk_factors['uncertainty_risk'] = min(float(jnp.std(samples)) / 20.0, 1.0)
        
        return risk_factors
    
    def batch_predict(self, 
                     demographics: List[Dict],
                     n_samples: int = 1000) -> List[LifeExpectancyPrediction]:
        """
        Batch prediction for multiple demographic profiles.
        
        Args:
            demographics: List of dicts with age, country, sex keys
            n_samples: Number of samples per prediction
            
        Returns:
            List of LifeExpectancyPrediction objects
        """
        predictions = []
        
        for demo in demographics:
            try:
                prediction = self.predict_life_expectancy(
                    age=demo['age'],
                    country=demo['country'],
                    sex=demo['sex'],
                    year=demo.get('year'),
                    n_samples=n_samples
                )
                predictions.append(prediction)
            except Exception as e:
                # Return default prediction for invalid demographics
                predictions.append(LifeExpectancyPrediction(
                    mean_life_expectancy=75.0,
                    confidence_interval=(65.0, 85.0),
                    uncertainty=10.0,
                    risk_factors={'error': 1.0}
                ))
        
        return predictions
    
    def get_model_info(self) -> Dict:
        """Get information about the trained model."""
        return {
            'n_mortality_records': len(self.mortality_data),
            'countries': self.graph_builder.countries,
            'age_range': (min(self.graph_builder.ages), max(self.graph_builder.ages)),
            'year_range': (min(self.graph_builder.years), max(self.graph_builder.years)),
            'n_nodes': len(self.graph.nodes),
            'n_edges': len(self.graph.edges),
            'n_factors': len(self.factors),
            'version': '0.1.1'
        }