A machine-learning approach to cardiovascular risk prediction in psoriatic arthritis

Machine-learning algorithms could outperform traditional methods in cardiovascular risk prediction among patients with PsA. Dear Sir, In patients with PsA an increased cardiovascular (CV) risk has been observed [1, 2]. Therefore, the identification of high CV risk PsA patients is crucial in order to implement preventive strategies. Although traditional CV risk calculators [such as Framingham Risk Score (FRS), Progetto CUORE and Systematic COronary Risk Evaluation (SCORE) risk scores] have already been assessed on PsA patients, they present numerous limitations, such as the underestimation of CV risk in PsA patients and poor calibration. Even the predictors correction indicated by EULAR does not result in substantial improvements [3, 4]. The inability to unravel complex relationships among heterogeneous variables in PsA patients could at least partially explain the poor performance of traditional CV risk algorithms in this population. Indeed, classical risk predictors are based upon the general assumption that each risk factor is linearly related to the CV disease outcome. Machine-learning (ML) techniques might be useful to overcome this limitation. ML is a subfield of artificial intelligence, recently employed in several clinical fields such as cardiology, that is able to learn how to solve a specific problem without being explicitly programmed for this. Nowadays, this is possible thanks to the huge quantity of data present in almost every field of science. ML's unique feature is that algorithms do not require many assumptions on the underlying data; therefore, non-linearities can be caught more easily, in order to face challenges that cannot be solved by traditional statistical methods [5]. Here, supervised ML techniques such as Support Vector Machine, Random Forest (RandF) and K-Nearest Neighbor were employed to predict CV risk in PsA patients. All the ML algorithms here employed were implemented in Python programming language. The results have been compared with those obtained by traditional CV risk predictors (such as FRS), discussed in depth in a recent work of our group [4]. Moreover, in this work the so called “features analysis”, based on random forest importances, was carried out to interpret the ML predictions and to better understand contributions of each single risk factor to the CVD outcome in PSA patients. To train the above-mentioned ML algorithms, a database of patients from the Framingham Heart Study (referred to as the general dataset) was employed. From the general dataset, an equal number of patients with and without a CV disease event were extracted to avoid bias in ML training. ML models were then validated on a PsA dataset containing 155 patients [4]. Concerning the PsA dataset, patients were Caucasian and lived in centre or south of Italy. Median (25th – 75th percentiles) age was 48 (40–55) years. During follow-up, 21 patients had a CV event (1.35 events per 100 patient-years): eight cases of myocardial infarction or unstable angina pectoris, three cases of stable angina pectoris, two cases of transient ischaemic attack, four cases of peripheral artery disease and four cases of heart failure. No fatal events were reported. This study complies with the Declaration of Helsinki and the locally appointed ethics committee (Università Campus Bio-Medico di Roma) has approved the research protocol. At this stage we limited the risk factors included in the ML models to: age, sex, systolic blood pressure, total cholesterol, smoking status and hypertension treatment, i.e. the same risk factors used in the Framingham Risk Score, which represent the classical risk predictors used as comparison terms in this work. Algorithms performances results are shown in Fig. 1A, in terms of sensitivity, specificity and accuracy. A comparison between FRS (10% cut-off EULAR version), Support Vector Machine (C = 0.1, γ = 0.01), RandF (splitting criterion = entropy, number of trees = 500) and K-Nearest Neighbor (K = 25) was performed. Algorithms performances results: a comparison between FRS and ML techniques (SVM, RF and KNN). (A) sensitivity, specificity and accuracy values, (B) ROC curves, (C) Random forest features importances. ML models exhibited a significantly higher sensitivity in PsA patients. Area under the curve values were: 0.7584 (95% CI 0.656, 0.85) for FRS, 0.8468 (95% CI 0.745, 0.925) for Support Vector Machine, 0.8346 (95% CI 0.732, 0.922) for RandF and 0.8010 (95% CI 0.705, 0.881) for K-Nearest Neighbor. Receiver operating characteristic curves are shown in Fig. 1B. Features analysis results are shown in Fig. 1C. In this case, RandF was pre-trained on the entire PsA dataset using all rheumatic features and it ranked those features based on their relative importance for a correct prediction. Interestingly, the results highlighted that FRS features are among the most important variables also in the PsA dataset (they cover about 50% of classification weight). Nevertheless, several PsA characteristics such as CRP and Psoriatic Area Severity Index were highlighted as relevant in the CV prediction (Fig. 1C). In conclusion, we showed that ML techniques could outperform traditional methods in terms of classification results in the field of CV risk prediction in rheumatic patients, and that they are promising techniques. However, it must be noted that a limitation of the present study is the PsA dataset’s size; it was too small to allow the training and validation of an ML algorithm specifically on PsA patients. Nevertheless, the size of the PsA dataset still allowed pre-training of ML algorithms and detection of the importance of risk factors to make effective CV predictions in PsA patients. Further studies should consider both training and validation on a larger PsA dataset, feeding the ML algorithm with specific PsA variables, such as CRP and Psoriatic Area Severity Index. Funding: No specific funding was received from any funding bodies in the public, commercial or not-for-profit sectors to carry out the work described in this manuscript. Disclosure statement: The authors have declared no conflicts of interest.