Dr Weng and colleagues reported improved accuracy of cardiovascular risk prediction when using machine-learning compared an established clinical algorithm with 8 variables recommended by American College of Cardiology Guidelines. Their 30 item-file failed to report some common conditions associated with stroke and/or myocardial infarction (including noncardiac surgery, human immunodeficiency virus, mental and professional stress, sleep deprivation, sedentarity), and did not report fitness (or results of stress tests), imaging (carotid artery intima-media thickness, calcic score…), ankle-brachial index, albuminuria, homocysteine or vitamin-D levels[1,2]. They did not collect some information that may play major role in cardiovascular risk including Mediterrean diet, as opposed to consumption of diet soda, and most of all red meat or trimethylamin-N-oxid (TMAO) levels[3,4]. Moreover, different antihypertensive therapy may not protect equally from cardiovascular consequences, despite similar blood pressure reduction.
Despite moderate improvement of up to 3.6% (“an encouraging step forward”), compared with a “simple clinical algorithm” comprising 8 variables, accuracy remained mild (improved from 0.728 to 0.745, .0760, and 0.764). I agree with the authors when they state, “These methods are not yet commonplace for developing risk prediction models in clinical datasets but their utility should be explored in future studies”. Accordingly, machine learning is certainly not ripe for clinical usage yet, as it is cumbersome, functions as a “black box”, and adds only minimal discrimination for now. I have no doubt we will find an algorithm that will more correctly identify subjects who could benefit from preventive measures, while avoiding unnecessary treatments in others.
It is interesting that missing values (i.e. body mass index) gave independent important prediction information, in the top 10 variables of some machine-learning algorithms. It was suggested that clinician may not record normal body mass index values, if patient is perceived at low cardiovascular risk. Another explanation may be that clinicians did not report even abnormal body mass index values, if patient is subjectively perceived at low cardiovascular risk (i.e. an overweighed fit and/or athlete subject). Even if subjective assessment of cardiovascular risk may be valuable-and is essential and irreplaceable in clinical practice, accuracy may be far less than reported by clinical and machine-learning algorithms. This being said, I cannot imagine today-and near future- clinical practice without machine helping clinicians to embrace the complexity of cardiovascular risk prediction, and subsequent preventive strategy (adapted from risk-adjusted algorithms).
Definitely, continued research on cardiovascular risk prediction with a major role of TMAO is eagerly needed for really improved algorithms (with accuracy over 0.8), in order to adapt preventive measures and decrease stroke, myocardial infarction and morbi-mortality in overall population.

References
[1] Lee D-C, Pate RR, Lavie CJ, Sui X, Church TS, Blair SN. Leisure-time running reduces all-cause and cardiovascular mortality risk. J Am Coll Cardiol 2014;64:472–81. doi:10.1016/j.jacc.2014.04.058.
[2] Beltrán LM, Rubio-Navarro A, Amaro-Villalobos JM, Egido J, García-Puig J, Moreno JA. Influence of immune activation and inflammatory response on cardiovascular risk associated with the human immunodeficiency virus. Vasc Health Risk Manag 2015;11:35–48. doi:10.2147/VHRM.S65885.
[3] Tang WHW, Wang Z, Levison BS, Koeth RA, Britt EB, Fu X, et al. Intestinal microbial metabolism of phosphatidylcholine and cardiovascular risk. N Engl J Med 2013;368:1575–84. doi:10.1056/NEJMoa1109400.
[4] Mitrou PN, Kipnis V, Thiébaut ACM, Reedy J, Subar AF, Wirfält E, et al. Mediterranean dietary pattern and prediction of all-cause mortality in a US population: results from the NIH-AARP Diet and Health Study. Arch Intern Med 2007;167:2461–8. doi:10.1001/archinte.167.22.2461.