Learning-by-doing is a fundamental concept in economics but a challenging one to document in high-skilled settings due to non-random assignment of workers to tasks and lacking performance measures. Our paper overcomes these challenges in the context of heart attack treatments in Sweden, where we exploit quasirandom assignment of physicians to patients. We document long learning curves, where physicians keep learning over the first 1000 treatments performed, affecting both proficiency and decision-making skills. These learning effects translate into effects on patient health, but only over the first 150 treatments performed, corresponding to one year of experience. Learning rates are higher for physicians who have worked with more experienced colleagues and who have gained more experience in treating complicated cases. Experienced physicians are more responsive to patient characteristics when deciding on treatments and experience from more recent heart attack treatments is more valuable than experience from more distant ones, suggesting that human capital depreciates. We also show that productivity growth keeps pace with wage growth over the first four years of the career but flattens out thereafter. Our results provide rare evidence on the existence of prolonged learning curves in high-skilled tasks and support the notion that learning-by-doing can be a powerful mechanism for productivity growth.