Deep learning (DL) is a specialized machine learning approach based on multiple-layered structures (algorithms) of artificial neurons, which are able to process information and learn by adjusting the weights at each synapse, enabling the performance of an intelligent task with high precision. Furthermore, these deep networks are able to ‘learn’ from large amounts of data and, similar to the human brain, can generalize concepts and apply these to new data with high accuracy. This ‘intelligence’ property has been expanded towards deep multimodal learning by integrating different sources of information (images, text, speech etc.) for further improving high-level multimodal decision support and reasoning (1). As a result, current DL systems are able to integrate and model heterogeneous data from an individual patient, across modalities and time, allowing better predictions and recommending treatment options tailored to each patient's individual characteristics and needs (2). This concept is illustrated in Fig. 1, highlighting the potential of DL to transform large-sized healthcare data into useful tools for advancing personalized medicine. DL algorithms are well-suited for large amounts of healthcare data and perform better than traditional statistical models at the expense of sacrificing interpretability for predictive power.

Accurate diagnosis is a milestone for personalized medicine. Ehteshami Bejnordi et al reported that 7 DL algorithms trained to detect metastases in hematoxylin- and eosin-stained tissue sections of lymph nodes of women with breast cancer, outperformed a panel of 11 pathologists with an area under the curve (AUC) of 0.994 (best algorithm) vs. 0.884 (best pathologist) (3). Madani et al presented a DL echocardiography system for the automated diagnosis of cardiac disease to address the issue of echocardiographic assessment inaccuracy (up to 30%) of echo reports, achieving an 80% accuracy in echocardiographic view classification and a 92.3% accuracy for left ventricular hypertrophy classification (4). Recently, Ding et al proposed a DL system demonstrating an improved early prediction of the final diagnosis of Alzheimer's disease (AD) (82% specificity at 100% sensitivity, on average of 75.8 months prior to the final diagnosis), utilizing ¹⁸F-fludeoxyglucose positron emission tomography (PET) of the brain, and therefore enhancing opportunities for early therapeutic interventions (5). Finally, another recent study presented a DL algorithm (CheXNeXt) that matched the expert human radiologist performance in diagnosing 11 lung pathologies (6).

DL has begun to revolutionize decision support systems in oncology due to the plethora of the available data and the heterogeneity of various malignancies, which require algorithms able to decipher hidden phenotype and genotype patterns, as well as their associations. Recently, Causey et al reported a DL system (NoduleX) which achieved an impressive accuracy (AUC of ~0.99) in nodule malignancy classification (>1,000 nodules) based on CT scans, matching the performance of experienced radiologists (7). Previously, Li et al reported a DL brain tumor radiogenomic system able to predict with high accuracy the IDH1 mutation status solely from multi-modal magnetic resonance imaging (MRI) data, achieving an impressive 95% AUC (8). Using a similar approach, Akkus et al demonstrated a DL system capable of predicting the codeletion of 1p/19q chromosome arms with high accuracy (87.7%) in 159 patients with low-grade glioma (9). In both cases, DL seems to enable the discovery of hidden imaging phenotype properties, which are associated with molecular/genetic information of high prognostic value.

The in silico prediction of therapy optimization is of utmost importance for personalized medicine. Bibault et al presented a DL radiomics system capable of predicting complete response following neo-adjuvant chemoradiation for locally advanced rectal cancer with the ultimate goal of identifying patients who would benefit more from conservative treatment rather than radical resections (10). From a bioinformatics standpoint, Yousefi et al developed the SurvivalNet framework based on DL and Bayesian optimization for predicting clinical outcomes from large amounts of data, which was generated by diverse genomic platforms (11). In a recent study on 768 patients, DL outperformed the widely used Cox proportional hazard regression model as well as other widely used models for survival analysis indicating that it can enhance the role of survival prediction in individualized clinical decision making (12). Reliable health predictors based on patient's heterogeneous electronic health records (EHRs) data integration is another important application field of DL. In this direction, Miotto et al developed the ‘Deep Patient’ prediction system trained on 700,000 EHRs and demonstrated high disease prediction performance for severe diabetes, schizophrenia and various malignancies (13).

Drug discovery for the personalization of therapy is another area that can benefit from the flexible architecture of DL systems towards de novo molecular design, reaction and bioactivity predictions (14). In this direction, IBM and Pfizer have joined forces for advancing immuno-oncology research with Watson for Drug Discovery (www.pfizer.com/news/press-release/press-releasedetail/ibm_and_pfizer_to_accelerate_immuno_oncology_research_with_watson_for_drug_discovery). Moreover, predicting environmental factors that influence health status is an integral part of preventive, personalized medicine. Ong et al proposed a time-series forecasting DL system for predicting the concentration of fine particulate matter PM2.5 (interfering with both human health and climate), in Japan, using exclusively real and publicly available sensor data, outperforming currently used methods (15).

DL architectures have already demonstrated significant technological advances across the personalized medicine informatics application spectrum, due to their versatility towards integrating multi-modal clinical data, discovering hidden properties and successfully generalizing results in unseen datasets. The number of DL applications for personalized medicine will continue to grow due to the high demand for such technologies and their anticipated socioeconomic impact. Although this process is facilitated by the widely available DL open computational frameworks and the abundance of large amounts of healthcare data, there is still skepticism regarding the clinical adoption of DL technologies emanating from the lack of causality and their ‘black-box’ nature. The inability to understand why an algorithm achieves generalization and performs so well, may be a critical factor inhibiting the clinical translation of DL technologies. To address this issue, several researchers are focusing on explaining the inner organization of DL networks. Shwartz-Ziv and Tishby, and Tishby and Zaslavsky have proposed an influential theory based on the ‘information bottleneck’ concept, according to which the network compresses input data as if by squeezing the information through a bottleneck, retaining only the features most relevant to the general concepts of the learning task (16,17). This compression process is pronounced at the higher layers of the network for promoting better predictions via preserving information relevant to the output labels at the expense of input data information, which is gradually ‘forgotten’. The confirmation of such theoretical explanatory models in conjunction with large-sized data validation across clinical sites and patient data registries, is a sine qua non condition for accelerating the clinical translation of DL personalized medicine decision support systems.