find_optimal_threshold.py

def find_optimal_threshold(y_true, y_scores, priority='sensitivity', min_value=0.95):
    """
    Find optimal threshold with constrained optimization.
    
    Parameters
    ----------
    y_true : array-like
        Ground truth labels (0 or 1)
    y_scores : array-like
        Predicted probabilities or continuous scores
    priority : str, 'sensitivity' or 'specificity'
        The metric to constrain (must be >= min_value)
    min_value : float
        Minimum acceptable value for the priority metric
    
    Returns
    -------
    dict with threshold, sensitivity, specificity, fpr
    """
    fpr, tpr, thresholds = roc_curve(y_true, y_scores)
    specificity = 1 - fpr
    
    if priority == 'sensitivity':
        constrained = tpr
        to_maximize = specificity
        constrained_name = 'sensitivity'
        maximize_name = 'specificity'
    elif priority == 'specificity':
        constrained = specificity
        to_maximize = tpr
        constrained_name = 'specificity'
        maximize_name = 'sensitivity'
    else:
        raise ValueError("priority must be 'sensitivity' or 'specificity'")
    
    valid_mask = constrained >= min_value
    
    if not valid_mask.any():
        print(f"Warning: No threshold achieves {constrained_name} >= {min_value}")
        print(f"Max achievable {constrained_name}: {constrained.max():.4f}")
        best_idx = np.argmax(constrained)
    else:
        valid_indices = np.where(valid_mask)[0]
        best_idx = valid_indices[np.argmax(to_maximize[valid_mask])]
    
    return {
        'threshold': thresholds[best_idx],
        'sensitivity': tpr[best_idx],
        'specificity': specificity[best_idx],
        'fpr': fpr[best_idx],
    }