To estimate the validity (sensitivity, specificity, and positive and negative predictive values) of a clinical algorithm as used by community health workers (CHWs) to detect and classify neonatal illness during routine household visits in rural Bangladesh. CHWs evaluated breastfeeding and symptoms and signs of illness in 395 neonates selected randomly from neonatal illness surveillance during household visits on postnatal days 0, 2, 5 and 8. Neonates classified with very severe disease (VSD) were referred to a community-based hospital. Within 12 hours of CHW assessments, physicians independently evaluated all neonates seen in a given day by one CHW, randomly chosen from among 36 project CHWs. Physicians recorded symptoms and signs of illness, classified the illness, and determined whether the newborn needed referral-level care at the hospital. Physicians' identification and classification were used as the gold standard in determining the validity of CHWs' identification of symptoms and signs of illness and its classification. CHWs' classification of VSD showed a sensitivity of 73%, a specificity of 98%, a positive predictive value of 57% and a negative predictive value of 99%. A maternal report of any feeding problem as ascertained by physician questioning was significantly associated (P < 0.001) with 'not sucking at all' and 'not attached at all' or 'not well attached' as determined clinically by CHWs during feeding assessment. CHWs identified with high validity the neonates with severe illness needing referral-level care. Home-based illness recognition and management, including referral of neonates with severe illness by CHWs, is a promising strategy for improving neonatal health and survival in low-resource developing country settings.