Automated Feature Engineering for Heart Failure

# Illustrative sketch (not from the paper)
import pandas as pd, torch

def pipeline(source_tables, target_features):
    # 1. Achievability
    unsupported = [f for f in target_features if f not in source_tables]
    # 2‑4. Standardize, deduplicate, aggregate per table
    summaries = []
    for name, df in source_tables.items():
        df = df.applymap(lambda x: None if pd.isna(x) or str(x).strip().lower() in ('', 'nan') else str(x).strip())
        df.columns = [c.upper() for c in df.columns]
        df = df.drop_duplicates(subset=['PATIENT_ID', 'EVENT_ID'])
        agg = df.groupby('PATIENT_ID').agg(event_cnt=('EVENT_ID', 'count'), first_ts=('TIMESTAMP', 'min'), last_ts=('TIMESTAMP', 'max')).reset_index()
        summaries.append(agg)
    # 5. Merge summaries
    from functools import reduce
    patient_df = reduce(lambda l, r: pd.merge(l, r, on='PATIENT_ID', how='outer'), summaries)
    # 6. Scoring rubric
    patient_df['COMPOSITE'] = patient_df['event_cnt'] * 0.1
    patient_df['EVIDENCE'] = 'rubric_v1'
    # 7. LLM audit (placeholder)
    patient_df['COMPOSITE'] = torch.tensor(patient_df['COMPOSITE'].values, dtype=torch.float32)
    # 8. Integrity check
    assert patient_df['COMPOSITE'].ge(0).all()
    return patient_df, unsupported

// Illustrative sketch (not from the paper)
const torch = require('torch-js'); // placeholder
function pipeline(sourceTables, targetFeatures) {
  const unsupported = targetFeatures.filter(f => !(f in sourceTables));
  const summaries = [];
  for (const [name, rows] of Object.entries(sourceTables)) {
    const agg = {};
    rows.forEach(r => {
      const pid = r.PATIENT_ID; const ts = r.TIMESTAMP;
      if (!agg[pid]) agg[pid] = { PATIENT_ID: pid, event_cnt: 0, first_ts: ts, last_ts: ts };
      agg[pid].event_cnt++; if (ts < agg[pid].first_ts) agg[pid].first_ts = ts; if (ts > agg[pid].last_ts) agg[pid].last_ts = ts;
    });
    summaries.push(Object.values(agg));
  }
  const merged = summaries.reduce((a, c) => { c.forEach(r => { a[r.PATIENT_ID] = Object.assign(a[r.PATIENT_ID] || { PATIENT_ID: r.PATIENT_ID }, r); }); return a; }, {});
  const patients = Object.values(merged);
  patients.forEach(p => { p.COMPOSITE = p.event_cnt * 0.1; p.EVIDENCE = 'rubric_v1'; p.COMPOSITE = torch.tensor([p.COMPOSITE]); });
  if (patients.some(p => p.COMPOSITE.dataSync()[0] < 0)) throw new Error('Integrity violation');
  return { patients, unsupported };
}