"""Current-visit predictors and rank-only AE screening for repeated patient splits.""" import json from pathlib import Path import numpy as np import pandas as pd from ampd.ae_latent_selection import centered_ranks,encode from ampd.ae_residual import FrozenMonthMLP,correct from ampd.protein_residual import branch_features from ampd.omics_boost import transform_pls METHODS=('reference','selected_ae','pca','pls') def arrays(path): with np.load(path,allow_pickle=False) as a: return dict(a) def rank_latents(z,samples,y): """Same earliest-patient Spearman ranking as screen_latents, without unused p-values.""" rows=[] for target in range(4): meta=samples.copy().reset_index(drop=True) meta['row']=np.arange(len(meta));meta['outcome']=y[:,4*target] first=meta[meta.outcome.notna()].sort_values(['uid','visit_month']).drop_duplicates('uid') ix=first.row.to_numpy(int);values=first.outcome.to_numpy(float) rho=centered_ranks(z[ix]).T@centered_ranks(values) for j in range(z.shape[1]): active=np.std(z[ix,j])>1e-10 and np.unique(values).size>=3 and len(ix)>=30 rows.append({'target':f'updrs_{target+1}','latent':j,'n_patients':len(ix),'rho':float(rho[j]) if active else np.nan}) return pd.DataFrame(rows) class SplitSeedPredictor: def __init__(self,bundle): bundle=Path(bundle) self.spec=json.loads((bundle/'model.json').read_text()) self.baseline=FrozenMonthMLP(bundle/'baseline') self.pca=arrays(bundle/'pca_extractor.npz');self.pca_head=arrays(bundle/'pca.npz') self.pls=arrays(bundle/'pls.npz') self.ae_pre=arrays(bundle/'ae_preprocessor.npz');self.ae_net=arrays(bundle/'ae_numeric.npz') self.ae_head=arrays(bundle/'selected_ae.npz') def predict(self,events): for e in events: if set(e)!={'visit_id','visit_month','proteins'}: raise ValueError('Only current visit_id, visit_month, proteins allowed') if len(e['proteins'])!=len(self.spec['proteins']): raise ValueError('Protein vocabulary mismatch') p=np.asarray([e['proteins'] for e in events],float);m=np.asarray([e['visit_month'] for e in events],float) if np.any(p<0):raise ValueError('Negative protein abundance') b=self.baseline.predict(m) z=encode(p,self.ae_pre,self.ae_net);z=(z-self.ae_head['mean'])/self.ae_head['scale'] return {'reference':b, 'selected_ae':correct(b,z@self.ae_head['coef'].T,self.ae_head['weights']), 'pca':correct(b,branch_features(m,p,'pca',self.pca)@self.pca_head['coef'].T,self.pca_head['weights']), 'pls':correct(b,transform_pls(p,self.pls)@self.pls['coef'].T,self.pls['weights'])}