"""Process features from arrived, label-free test events only. This module has no clinical-table input or cohort lookup. Learned model and preprocessing parameters may be loaded separately by the inference process. """ from __future__ import annotations import numpy as np class StreamProcessFeatures: FIELDS = {"visit_id", "patient_id", "visit_month"} def __init__(self, months, mode="observed"): if mode not in {"observed", "current_only"}: raise ValueError(mode) self.months = [int(m) for m in months] self.mode = mode self.history = {} self.last_month = -1 self.seen_visits = set() @property def names(self): return [f"v{m}" for m in self.months] + ["control", "gap", "prior_count", "visit_month"] def update(self, event): if set(event) != self.FIELDS: raise ValueError("Inference events must contain exactly visit_id, patient_id, visit_month") patient, month = int(event["patient_id"]), int(event["visit_month"]) visit = str(event["visit_id"]) if month < self.last_month: raise ValueError("Test events must arrive in chronological order") if visit in self.seen_visits: raise ValueError("Duplicate test event") self.last_month = month self.seen_visits.add(visit) if self.mode == "current_only": flags = [1.0 if m == month else -1.0 for m in self.months] return np.asarray(flags + [-1., -1., -1., float(month)], dtype=np.float32) previous = self.history.get(patient, set()) past = sorted(m for m in previous if m < month) available = previous | {month} self.history[patient] = available flags = [-1. if m > month else float(m in available) for m in self.months] first_nonzero = min((m for m in available if m > 0), default=None) control = -1. if month < 12 else float(first_nonzero == 12) gap = float(month - past[-1]) if past else -1. return np.asarray(flags + [control, gap, float(len(past)), float(month)], dtype=np.float32) def features_for_events(events, months, mode="observed"): builder = StreamProcessFeatures(months, mode) rows, visits = [], [] for event in events: rows.append(builder.update(event)) visits.append(event["visit_id"]) return np.asarray(rows, dtype=np.float32), visits, builder.names