Just as the map is not the territory, so too an algorithm is never the care that is given. Algorithms, neural networks, guidelines, and protocols—these all can only ever model aspects of a reality that is always more complex and fickle. As we move into a world dominated by algorithms and machine-learned clinical approaches, we must deeply understand the difference between what a machine says and what we must do.
A growing number of research papers are reporting impressive diagnostic performance for computer systems built using machine learning. Deep learning techniques in particular are transforming our ability to interpret imaging data.2 In The Lancet Oncology, Xiangchun Li and colleagues3 report a retrospective preclinical study applying deep learning and statistical methods to diagnose thyroid cancer using sonographic images. Their results are impressive. When compared with six radiologists on unseen data, in an internal validation dataset, the system correctly detected about the same number of cancers (sensitivity 93·4% [95% CI 89·6–96·1] with the algorithm vs 96·9% [93·9–98·6] with the radiologists; p=0·003) but had far fewer false-positives (specificity 86·1% [95% CI 81·1–90·2] with the algorithm vs 59·4% [53·0–65·6] with the radiologists; p<0·0001).