Artificial Intelligence in Clinical Practice: How AI Technologies Impact Medical Research and Clinics

Artificial Intelligence in Clinical Practice: How AI Technologies Impact Medical Research and Clinics compiles current research on Artificial Intelligence within medical subspecialties, helping practitioners with diagnosis, clinical decision-making, disease prediction, prevention, and the facilitation of precision medicine. The book defines the basic concepts of big data and AI in medicine and highlights current applications, challenges, ethical issues, and biases. Each chapter discusses AI applied to a specific medical subspecialty, including primary care, preventive medicine, general internal medicine, radiology, pathology, infectious disease, gastroenterology, cardiology, hematology, oncology, dermatology, ophthalmology, mental health, neurology, pulmonary, critical care, rheumatology, surgery, and OB-GYN.

This is a valuable resource for clinicians, students, researchers and members of medical and biomedical fields who are interested in learning more about artificial intelligence technologies and their applications in medicine.

• Provides the history and overview of the various modalities of AI and their applications within each field of medicine
• Discusses current AI-based medical research, including landmark trials within each field of medicine
• Addresses the current knowledge gaps that clinicians commonly face that prevent the application of AI-based research to clinical practice
• Encompasses examples of specific cases and discusses challenges and biases associated with AI

• Language: English
• Publisher: Academic Press
• Release date: Sep 13, 2023
• ISBN: 9780443156892

Book preview

Chapter 1

Artificial intelligence in primary care

Adham El Sherbini¹, Benjamin S. Glicksberg² and Chayakrit Krittanawong³,    ¹Faculty of Health Sciences, Queen’s University, Kingston, ON, Canada,    ²Hasso Plattner Institute for Digital Health, Icahn School of Medicine at Mount Sinai, New York, NY, United States,    ³Cardiology Division, NYU Langone Health and NYU School of Medicine, New York, NY, United States

Abstract

An overview of the promise, restrictions, and difficulties of artificial intelligence (AI) in revolutionizing primary care is given in this chapter. The application of AI, machine learning, and deep learning in healthcare has made significant strides in several primary care domains, including screening, preoperative care, and disease diagnosis. Numerous diseases, including cancer, cardiovascular disease, sexually transmitted diseases, osteoporosis, diabetes mellitus, hypertension, ophthalmic diseases, and obstructive sleep apnea, have been successfully predicted, diagnosed, and risk assessed using AI models. For the widespread adoption of AI in primary care, issues including data quality, interpretability, and regulatory constraints must be resolved. Nevertheless, by enhancing disease diagnosis, risk assessment, and personalized care, AI has the potential to transform primary care.

Keywords

Artificial intelligence; machine learning; primary care; screening; preoperative care; risk assessment

Introduction

The collaboration of artificial intelligence (AI) and primary care to blossom a captivating new field of study is worth understanding. How the subsets of AI, machine learning (ML), and even deeper, deep learning (DL) interact with the countless facets of healthcare is significant for a progressive future. Over the course of a half-century, AI has made tremendous strides in many areas of high-value primary care clinical practice (e.g., annual physical screening, preoperative risk stratification, vaccination), as will be discussed in this chapter. Nonetheless, there are limitations and challenges that have prolonged its implementation. This chapter briefly covers AI’s potential, limitations, and challenges in primary care transformation.

Potentials

In nearly all, research has been conducted to evaluate AI’s potential within its implementation. Although certain sectors of healthcare are more promising than others, ML typically provides a unique outlook on healthcare practices and opens doors for new strategies to be tested. To date, and speaking to its ever-lasting effects, AI’s applications in primary care can be broken down into three sections: screening, preoperative management, and detection. In all three, AI has made progressive strides, as seen through research studies and innovative healthcare start-ups (Fig. 1.1).

Figure 1.1 Barriers, current, and future direction of artificial intelligence in primary care.

Screening

Regarding screening, ML models have been applied to predict, diagnose, and assess the risk of cancer, cardiovascular diseases (CVD), sexually transmitted diseases (STDs), osteoporosis/osteomalacia, diabetes mellitus (DM), hypertension, hyperlipidemia, ocular diseases, and obstructive sleep apnea (OSA).

Cancer screening

Cancer screening measures have been relatively positive with the introduction of ML into breast, lung, prostate, colorectal, and cervical cancer. Certain ML algorithms, such as CervDetect [composed of Pearson correlation, random forest (RF), and shallow neural networks (NNs)] and DeepCervix (DL-based framework), are independently developed and assessed on cervical cancer screening and have achieved exceptionally high accuracies (>90%) [1–3]. Similarly, numerous studies have developed or compared a number of analytical models (ML and DL) in the screening and prediction of lung cancer [4,5]. The general conclusion is that most of these models achieve high accuracy and outperform traditional or previously established models, such as the 2012 Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial risk mode (mPLCOm2012) [6]. Physicians utilize several types of diagnostic modalities, such as electronic health records (EHRs), computed tomography (CT) scans, and low-dose CT scans to train these models to predict lung cancer. Comparatively, mammograms are the focal images utilized for training models to screen for breast cancer [7,8]. Breast cancer-focused studies presented similar results, in which AI models had great accuracy in the screening for breast cancer and even outperformed the established BOAIDICEA model [8]. As for colorectal cancer, fecal microbiotas and additional modalities were utilized to compare numerous models predicting colorectal cancer, and models typically had higher accuracy (>90%) [9,10]. ML has also been utilized in chemotherapy for pretreatment prediction. In one study, multiple ML models [penalized linear regression, support vector machine (SVM), and RF] were utilized to predict left ventricular dysfunction in lymphoma or breast cancer patients. The RF was the best performing, with an accuracy of 0.94, a sensitivity of 0.81, and a specificity of 0.98. Similarly, prostate biopsies, EHRs, and PSA levels were applied to countless ML models to predict prostate cancer, and all studies concluded that the AI was effective in its screening [11,12].

Cardiovascular disease predictions or surveillance

Studies have evaluated AI applications on cardiovascular risk predictions (e.g., hypertension, CVD, coronary calcium scoring) through a number of ML models [RF, logistic regression (LR), gradient boosting (GB), NN, convolutional neural network (CNN), long short-term memory, etc.] [13–15]. Generally, ML models are high performing and have outperformed traditional scores, such as the Framingham score and a Cox PH model [16]. Similar findings are concluded in hypertension, resistant hypertension (defined as uncontrolled blood pressure (BP) despite the use of ≥5 antihypertensive medication, including diuretics), refractory hypertension (defined as uncontrolled BP despite the use of ≥5 antihypertensive medication, including diuretics and mineralocorticoid receptor antagonist), and estimate BPs, where models have reported area under the curve (AUC) scores as high as 0.92 [17,18]. Wearable technology or mobile health device with ML could be used in ambulatory BP monitoring and lifestyle recommendations. Given the significance of ambulatory monitoring for cardiovascular management, the introduction of novel monitoring devices can be effective given ML’s assistance to increase accuracy and actionability. To date, the combination of portable sensors alongside ML models allows for near-real-time diagnosis [19]. Specifically, wearable sensors detecting relevant biosignals can be interpreted through ML algorithms for heightened accuracy. Similarly, the combination of wearable technology and ML can be utilized for personalized care, such as improvements in BP. In Chiang et al., RF with Shapley-Value-based Feature Selection was used to apply a highly accurate and personalized BP model [20]. Comparatively, a systematic review was conducted on ML techniques combined with smartphones for the monitoring of BP [21]. ANN and SVMs were the most commonly used input features ranging from five to 233 [21].

AI research also allows for the identification of significant clinical factors, such as hypertension research, where they found educational status, tobacco use, mental health, socioeconomic status, lifestyle (e.g., consumption of fruits and vegetables, physical activity), occupation, age, gender, history of diabetes, abdominal obesity, and family history (e.g., mothers history of hypertension, and history of high cholesterol) [22]. Comparatively, when investigating DM, ML models have outputted very similar results [23–25]. A number of approaches have been utilized for training the models, such as cardiorespiratory fitness records and administrative data, which assisted in outperforming established models and detecting adverse complications [22–25]. Coronary calcium scoring, a noninvasive tool for risk assessment of coronary artery disease, can utilize AI to enhance its clinical applications. Eng et al. developed two DL models capable of automating coronary artery calcium (CAC) scoring [26]. One model, a gated coronary CT model intended for CAC scoring, presented with a significantly high agreement (mean difference in scores=−2.86; Cohen’s kappa=0.89, P<0.0001) [26]. The other model, a DL algorithm, was trained on the MESA study to undergo CAC scoring [26]. Across all levels of CAC, sensitivity ranged from 80%–100%, and PPV ranged from 87%–100% [26]. A similar study by van Velzen et al. trained and validated. DL model on CT examinations for automatic calcium scoring [27]. The DL model’s intraclass correlation coefficients for CAC are 0.79–0.97 [27]. Similarly, Mu et al. trained and validated a DL model to quantify CAC scores from a singular coronary CT angiography (CTA) [28]. The risk categorization agreement between DL CTA and noncontrast CT CAC scores was exceptional (weight k=0.94) [95% confidence interval (CI): 0.91, 0.97] [28]. To add, the positive correlation between the semiautomatic noncontrast CT CAC score and the automatic DL CTA was considered excellent (Pearson correlation=0.96; r²=0.92) [28]. AI may also have the potential to conduct risk prediction of CAC scores from retinal photographs, as observed in Rim et al. [29]. Five datasets from Singapore, the United Kingdom, and South Korea were used to train and validate DL models’ ability to predict CAC (RetiCAC) [29]. RetiCAC was performed between all single clinical parameter models, and an AUROC of 0.742 was achieved [29]. In Siva Kumar et al., the prediction of ASCVD through an ML-centric ECG risk score in combination with or without CAC scores was measured [30]. The Mecg had a strong association and discrimination of MACE (C-index: 0.7), whereas the Mecg+cac was associated with MACE (C-index: 0.71) [30].

In general, the pooled cohort equations (PCE) could predict patients’ atherosclerotic cardiovascular disease risk [31]. However, PCE performance may over- or underestimate in certain ethnicity, and PCE performance among Asians or Hispanics remain inconclusive [31]. ML could potentially improve risk prediction in certain ethnic population [31]. In Ward et al., ML models (LR, GBM, and extreme GB) were trained on EHR for the ACSVD risk assessment of multiethnic patients [32]. GBM was well performing in the cohort and achieved an AUC score of 0.835, which was better than the PCE (AUC 0.775) [32]. Integration of additional EHR data had minimal effect on GBM’s performance (before: AUC 0.784, after AUC 0.790) [32].

Another CVD tool, the American College of Cardiology/American Heart Association (ACC/AHA) Pooled Cohort Equations Risk Calculator, may possibly underestimate the risk of atherosclerotic CVD in specific individuals [33]. With that, Kakadiaris et al., set to compare the ACC/AHA CVD Risk Calculator to an SVM-based ML Risk Calculator using the Multi-Ethnic Study of Atherosclerosis (MESA) [33]. The ML Risk Calculator had AUC, sensitivity, and specificity of 0.92, 0.86, and 0.95 [33]. This was better performing than the ACC/AHA CVD Risk Calculator, which resulted in an AUC of 0.71, a sensitivity of 0.76, and a specificity of 0.56 [33]. Although more studies are required to validate these findings, the ML Risk Calculator recommended fewer drug therapies but did miss fewer events [33]. Notably, the combination of an ML model with the established CAC score evoked a high AUC score in the prediction of coronary artery disease (CAD) [30–32,34–36]

In Petrazzini et al., an ML algorithm was developed based on clinical features extracted from HER for the prediction of CAD [37]. The algorithm was applied in the UK Biobank (population-based cohort) and BioMe (multiethnic clinical care cohort) [37]. Relative to PCE, the ML model improved CAD prediction in UK Biobank and BioMe by 9 and 12%, respectively [37]. Similarly, Cho et al. evaluated preexisting cardiovascular risk prediction models while developing ML-centric algorithms on a healthy Korean population [38]. Overall, the preexisting models were deemed moderate to good, with PCE having a C-statistic of 0.738 [38]. Regarding the ML models, an NN model produced the greatest C-statistic of 0.751 [38].

Sexually transmitted disease screening

As for STDs, studies have been conducted in rural areas of South Africa, Uganda, and Kenya with reported outcomes on human immunodeficiency virus (HIV), gonorrhea, chlamydia, and syphilis [39,40]. Overall, studies have been generally positive, with high AUC scores across a number of models and diagnostic training tools. For instance, in Turbe et al., DL was trained and validated on rapid HIV tests to predict them as positive or negative [39]. The model achieved high specificity (100%) and sensitivity (97.8%) [39]. Similarly, in Balzer et al., ML was applied to detect individuals at high risk of HIV across Rural Kenya and Uganda [40]. Regarding efficiency, for a fixed sensitivity of 50%, the model-based strategy (based on LR) targeted 27%, whereas the ML targeted only 18%, and it improved sensitivity [40].

Ocular diseases

Similarly, the assessment of ocular diseases, such as glaucoma, age-related macular degeneration, and retinopathy, have been promising [41–43]. Models have shown exceptionally high accuracy scores, especially when trained on retinal images. Additionally, using retinal fundus images has allowed models to surmount eye diseases and precisely predict cardiovascular risk, age, sex, systolic BP, adverse cardiac events, and smoking status. The unmatched ML models have also shown progress in CAD, OSA, and osteoporosis through CTs, questionaries, and EHRs [41–43].

Osteoporosis screening

AI has also been assessed in osteoporosis screening through several diagnostic modalities. For example, in Sebro et al., SVMs were trained on the wrist and forearm CT scans of 196 patients [44]. Radial-basis-function kernel SVM was the best-performing model for the prediction of osteoporosis (AUC=0.818) [44]. Similarly, Hsieh et al., assessed the risk of fracture and bone mineral density through training in DL on plain radiographs [36]. The model was high performing for spine osteoporosis, hip osteoporosis, high hip fracture risk, and high 10-year major fracture risk, with accuracies of 86.2%, 91.7%, 90.0%, and 95.0%, respectively [36]. In Zhang et al., another diagnostic, lumbar spine X-ray images, were utilized for training a deep convolutional NN model in screening for osteopenia and osteoporosis [45]. The model was trained on 1616 X-rays and tested on two test datasets (396 and 348 images) and a validation dataset (204 images) [45]. This model may have the potential to screen for osteoporosis (AUC 0.767) and osteopenia (AUC 0.787) [45]. Comparatively, another study Lim et al., utilized ML for the screening of femoral osteoporosis based on abdomen-pelvic CT and extracted radiomic features [46]. The ML model achieved accuracies of 92.9% and 92.7% in training and validation datasets, respectively. AUC scores of 95.9% and 96.0% in training and validation datasets were also reported [46]. Similar findings were reported by Pickhardt et al. where a DL tool was evaluated for its osteoporosis prediction based on CT images [47]. In regard to success rate, the DL model outperformed the traditional image processing BMD algorithm (99.3% vs 89.4%) [47]. AI has the potential to screen for osteoporosis through several modalities and outperform established models [47].

Obstructive sleep apnea screening

Given the increasing prevalence of OSA, preliminary studies evaluating the potential for AI to screen OSA have been noted. For one, in Alvarez et al., the assessment of the diagnostic utility of airflow recordings in combination with at-home oximetry through ML was evaluated [48]. Measurements are captured through at-home polysomnography, and regression SVMs were utilized to predict the apnea-hypopnea index [48]. When combining the two metrics, the model was better performing (kapaa:0.71; 4-class accuracy: 81.3%) than airflow (kappa: 0.42, 4-class accuracy: 61.5%) and oximetry (kappa: 0.61; 4-class accuracy: 75.0%) [48]. Similar findings were observed in Stretch et al., where ML models were trained on home sleep apnea testing for OSA [49]. All ML algorithms prompted higher partial AUPRC scores than LR (0.574), and the RF model was the best-perform (pAUPRC 0.862) [49]. Comparatively, in Kelly et al., in-home polysomnography was compared to the mandibular movement in combination with ML regarding the diagnosis of OSA [50]. Forty patients underwent both diagnostics, and there was good agreement between the two tools (median bias 0.00; 95% CI −23.25 to +9.73 events/hour) [50]. On a broader scale, Ferreira-Santos et al. conducted a systematic review on early OSA screening through ML [51]. Of 63 included studies, 23 studies focused on diagnostic development, and 26 studies included internal validation [51]. Studies included a wide range of ML algorithms, including LR, SVM, linear regression, NN, and DT, and RF, to name a few [51]. More notably, the highest reported AUROC was 0.98 (0.96–0.99). That said, there remains an absence in external validation within larger cohorts [51].

Preoperative risk assessment

Another sector of primary care where AI has asserted grounds is preoperative management, where pre- and intraoperative data are utilized for training ML models to predict postoperative mortality, postoperative complications, and prolonged length of stay in the intensive care unit (ICU) [52–54]. Functional capacity is the most important part of preoperative risk assessment for cardiovascular complications [55]. ML has evaluated functional capacity through intelligence robotics and wearable monitoring. Intelligent robotic systems may improve functional capacity evaluations in combination with ML methods [56]. ML allows for the fusion of expert human intelligence and robotic systems, which can allow for increased quantification [56]. As for wearable technology, several studies have used this ML-centric approach for the monitoring of rehabilitation. In Canniere et al., cardiac rehabilitation patients were equipped with a multiparameter sensor throughout activity for the interpretation of functional capacity [57]. In combination with chronotropic response and effort, SVM presented a mean absolute error of 42.8 m (±36.8 m) [57]. In Rens et al., CVD patients underwent a home-based 6-minute walk test (6MWT) which was evaluated by LR with forward feature selection for prediction of functional capacity [58]. The model could evaluate frailty with high specificity (85%) and sensitivity (90%) [58]. Another study used wearable monitoring devices for the prediction of metabolic equivalents (METs) [59]. Five multiple-regression models were used, and all presented significant improvement in the mean absolute percentage error of METs within the high-intensity group [59]. Similarly, Fuller et al. evaluated how wearable devices (Apple Watch Series 2, Fitbit Charge HR2, and iPhone 6S) could elevate ML to detect lying, sitting, walking, and running [60]. Rotation Forest models produced the highest accuracies for Fitbit and Apple Watch (90.8 and 82.6%, respectively) [60]. Overall, ML and wearable technology integration to evaluate functional capacity could be used in preoperative risk assessment.

The Revised Cardiac Risk Index (RCRI) is also recommended to use in preoperative risk assessment to predict perioperative cardiac complications. However, RCRI did not perform well at predicting cardiac events after vascular noncardiac surgery or at predicting death [61]. Another study showed that the RCRI score failed to accurately predict the risk of cardiac complications in patients undergoing elective resection of lung cancer [62]. Hofer et al. evaluated the ability of a rules-based algorithm to identify patients with diseases from the RCRI [63]. The algorithm produced a higher incidence of definite or likely disease than that of the anesthesiologist [63]. The changed RCRI from the algorithms achieved an AUROC score of 0.70 (0.67–0.73) in predicting in-hospital mortality and took 12.64±1.20 to go through 1.4 million patients [63].

Another ML-based preoperative risk assessment metric is the Cardia Comorbidity Risk Score (CCoR), which screens for major adverse cardiac events (MACE). This is typically conducted following elective knee and hip arthroplasty. The tool was well-performing for those with no history of RCRI conditions and even out-performed RCRI in detecting patients undergoing knee and hip arthroplasty [64] at a high risk of MACE [65]. The risk score utilizes ML to predict MACE through EHRs [65]. In Onishchenko et al. cohort, the CCoR achieved an AUROC score of 80.1% and 80.0% for males and women, respectively [65].

Notable studies have been conducted on countless surgeries in several populations, including those infected with SARS-CoV-2. The COVIDSurg Cohort Study was an international multicentered cohort of patients diagnosed with SARS-CoV-2 either 30 days postsurgical operation or seven days presurgical operation [64]. Through linear and nonlinear modeling, five relevant features were combined in 26 predictor sets that were fitted with either LR, DT, or RF [64]. The COVIDSurg Mortality Score achieved AUROC scores of 0.73 and 0.80 in the discrimination and validation sets, respectively [64]. This was comparable to that of previously published nonsurgical COVID risk assessments [64]. In another study, Chiew et al. compared ML models to the Combined Assessment of Risk Encountered in Surgery model and the American Society of Anaesthesiologists-Physical Status in the prediction of 30-day postoperative mortality and ICU stay longer than 24 hours [66]. The ML models (RF, adaptive boosting (AB), GB, and SVM), which were trained on EHR data, prompted high AUROCs but low sensitivities [66]. Specifically, GB was the highest-performing model with AUPRC scores of 0.23 and 0.38 for postoperative mortality and ICU admissions, respectively [66]. It was noted that the ML models predicted all negatives within the dataset, which was mostly negative [66]. In another study, Sahara et al. concluded that an ML algorithm outperformed the established American College of Surgeons National Surgery Quality Improvement Program surgical risk calculator in predicting 30-day mortality in hepatopancreatic surgical candidates [67]. Regardless of the training tools, nearly all studies have performed exceptionally accurately [67]. Such research is instrumental as it allows for forecasting clinical decisions and primary care resources, such as ICU beds and additional costs. It also can allow for cost-effective analysis of the patients and their providers.

Vaccination

ML may also have the potential to aid in prioritizing, screening and increasing adherence to vaccination. For instance, in Couto et al., ML was utilized to determine priority vaccination groups to reduce mortality rates. A GB model was trained on health variables (sex, age, and chronic health conditions) on a cohort of individuals hospitalized in one of 336 Brazilian hospitals [68]. Based on in-hospital death, the GB models achieved an AUROC score of 0.80 [68]. Regarding other vaccines, Hada et al. set out to develop a hybrid ML model to recommend vaccines for patients based on host-based factors (age, sex, medical history, postvaccination recovery rate, mortality rate, and symptoms). Sixteen experiments were conducted on COVID-19 and FLU3 vaccines, and there were clear differences ins cores based on the fore-mentioned factors. In regards to adherence, Kim et al., evaluated the potential of ML models to detect Korean adult CVD patients with low adherence to the influenza vaccination [69]. A total of 815 adults from the Fifth Korea National Health and Nutrition Examination Survey (KNHANES V) were utilized to assess several ML algorithms (LR, RF, SVM, XGB) [69]. The dataset was divided into <65 and ≥65 years old, as the elderly were provided with free immunization. For the ≥65 years old cohort, SVM had the highest accuracy (68.4%), and in the <65-year-old group, XGB and RF had the highest accuracies (84.7%) [69].

Detection

The final central part of primary care that ML has exceptionally investigated is the detection of COVID-19, flu, influenza, chronic coughs, pharyngitis, and atrial fibrillation (AF) [70,71]. Research has garnered rather promising results for all the listed conditions, as AUC scores are at an all-time high. For example, in Zhou et al., an XGBoost algorithm was utilized to differentiate cases of influenzas and COVID-19. The model achieved an AUC score of 0.85 in the external validation and 0.93 in the testing dataset. Similar findings follow through for all previously listed conditions. Another noteworthy mention is in the detection of AF, where ML models installed onto wearable devices can be trained on the health records, waveform data, RR intervals, W-PPG, and pulse oximetry data to accurately detect AF [72,73]. In Perez et al., a clinical study was conducted utilizing Apple Watch optical sensors were utilized to detect AF [74]. In the study, if participants had an irregular pulse that indicated possible AF, they were mailed an electrocardiography (ECG) patch [74]. Of those that returned an ECG (n=450), 34% of participants had AF [74].

Limitations and challenges

While favorable outcomes have been discussed regarding its applications, no randomized clinical trials are available. Indeed, AI suffers from several limitations and drawbacks that are difficult to address and resolve [75–77]. First, the introduction of AI has concurrently introduced a number of issues that were not previously an obstacle. For example, an escalation in the lack of trust by patients toward their healthcare services can arise. As bedside manners and trust are key elements in medical education and patient interaction, reliance on technology could break that trust. Frontiers in AI research and its applications should focus on developing a human interface that is capable of providing qualities that typical healthcare practitioners would otherwise deliver. Second, education on technology and AI would have to be implemented into medical education for upcoming and ongoing healthcare workers. This would incur additional costs and extensive time, as it is not a primary focus of these employees. Third, the implementation of AI into the workplace would be exceptionally costly, and a cost-effective analysis of all aspects of its application would be required. Fifth, a lack of accountability persists in primary care once AI is utilized. Typically, a wrong diagnosis or risk assessment would fall in the hands of the physician or healthcare provider in charge. If this specific task is to be replaced by AI, there serves a lack of responsibility. Even if the physician and AI are utilized for the same assessment, healthcare workers will rely on the model and question their individual opinions. This raises numerous legal and ethical concerns as the absence of accountability is a predicament. To be addressed, legal identification of ownership over the algorithms and healthcare practices should be strictly identified. Sixth, there are several limitations to AI in clinical research. For one, many studies suffer from a lack of external validation, making the findings questionable. Lastly, additional limitations that are quite common include a single center, retrospective data, missing data, and a lack of comparison to traditional scores. Although these drawbacks are more difficult to address, it’s necessary to acknowledge them when interpreting the results and conclusions of studies.

Conclusion

AI applications in primary care, though early, are up-and-coming. Study after study, ML models typically outperform established scores or formulas in cancer screening, preoperative risk prediction, vaccination guidance, detection, and prediction of nearly all aspects of primary care. However, its implementation in day-to-day primary care is still in its infancy. Large prospective studies and randomized clinical trials focusing on remote home monitoring and lifestyle monitoring are needed.

References

1. Mehmood M, Rizwan M, Gregus ml M, Abbas S. Machine learning assisted cervical cancer detection. Front Public Health. 2021;9:788376 https://doi.org/10.3389/fpubh.2021.788376.

2. Rahaman MM, Li C, Yao Y, et al. DeepCervix: a deep learning-based framework for the classification of cervical cells using hybrid deep feature fusion techniques. Comput Biol Med. 2021;136:104649 https://doi.org/10.1016/j.compbiomed.2021.104649.

3. Hou X, Shen G, Zhou L, Li Y, Wang T, Ma X. Artificial intelligence in cervical cancer screening and diagnosis. Front Oncol. 2022;12:851367 https://doi.org/10.3389/fonc.2022.851367.

4. Huang B, Sollee J, Luo YH, et al. Prediction of lung malignancy progression and survival with machine learning based on pre-treatment FDG-PET/CT. eBioMedicine. 2022;82:104127 https://doi.org/10.1016/j.ebiom.2022.104127.

5. Wang S, Zhang H, Liu Z, Liu Y. A novel deep learning method to predict lung cancer long-term survival with biological knowledge incorporated gene expression images and clinical data. Front Genet. 2022;13:800853 https://doi.org/10.3389/fgene.2022.800853.

6. Gould MK, Huang BZ, Tammemagi MC, Kinar Y, Shiff R. Machine learning for early lung cancer identification using routine clinical and laboratory data. Am J Respir Crit Care Med. 2021;204(4):445–453 https://doi.org/10.1164/rccm.202007-2791OC.

7. Cardoso MJ, Houssami N, Pozzi G, Séroussi B. Artificial intelligence (AI) in breast cancer care—leveraging multidisciplinary skills to improve care. Breast. 2021;56:110–113 https://doi.org/10.1016/j.breast.2020.11.012.

8. Ming C, Viassolo V, Probst-Hensch N, Dinov ID, Chappuis PO, Katapodi MC. Machine learning-based lifetime breast cancer risk reclassification compared with the BOADICEA model: impact on screening recommendations. Br J Cancer. 2020;123(5):860–867 https://doi.org/10.1038/s41416-020-0937-0.

9. Ho C, Zhao Z, Chen XF, et al. A promising deep learning-assistive algorithm for histopathological screening of colorectal cancer. Sci Rep. 2022;12(1):2222 https://doi.org/10.1038/s41598-022-06264-x.

10. Wang KS, Yu G, Xu C, et al. Accurate diagnosis of colorectal cancer based on histopathology images using artificial intelligence. BMC Med. 2021;19(1):76 https://doi.org/10.1186/s12916-021-01942-5.

11. Bashashati A, Goldenberg SL. AI for prostate cancer diagnosis—hype or today’s reality?. Nat Rev Urol. 2022;19(5):261–262 https://doi.org/10.1038/s41585-022-00583-4.

12. Janowczyk A, Leo P, Rubin MA. Clinical deployment of AI for prostate cancer diagnosis. Lancet Digit Health. 2020;2(8):e383–e384 https://doi.org/10.1016/S2589-7500(20)30163-1.

13. Adedinsewo DA, Pollak AW, Phillips SD, et al. Cardiovascular disease screening in women: leveraging artificial intelligence and digital tools. Circ Res. 2022;130(4):673–690 https://doi.org/10.1161/CIRCRESAHA.121.319876.

14. Aryal S, Alimadadi A, Manandhar I, Joe B, Cheng X. Machine learning strategy for gut microbiome-based diagnostic screening of cardiovascular disease. Hypertension. 2020;76(5):1555–1562 https://doi.org/10.1161/HYPERTENSIONAHA.120.15885.

15. Lou YS, Lin CS, Fang WH, et al. Artificial intelligence-enabled electrocardiogram estimates left atrium enlargement as a predictor of future cardiovascular disease. J Pers Med. 2022;12(2):315 https://doi.org/10.3390/jpm12020315.

16. Alaa AM, Bolton T, Di Angelantonio E, Rudd JHF, van der Schaar M. Cardiovascular disease risk prediction using automated machine learning: a prospective study of 423,604 UK Biobank participants. PLoS One. 2019;14(5):e0213653 https://doi.org/10.1371/journal.pone.0213653.

17. Chaikijurajai T, Laffin LJ, Tang WHW. Artificial intelligence and hypertension: recent advances and future outlook. Am J Hypertens. 2020;33(11):967–974 https://doi.org/10.1093/ajh/hpaa102.

18. Hung MH, Shih LC, Wang YC, et al. Prediction of masked hypertension and masked uncontrolled hypertension using machine learning. Front Cardiovasc Med. 2021;8:778306 https://doi.org/10.3389/fcvm.2021.778306.

19. Krittanawong C, Rogers AJ, Johnson KW, et al. Integration of novel monitoring devices with machine learning technology for scalable cardiovascular management. Nat Rev Cardiol. 2021;18(2):75–91 https://doi.org/10.1038/s41569-020-00445-9.

20. Chiang PH, Wong M, Dey S. Using wearables and machine learning to enable personalized lifestyle recommendations to improve blood pressure. IEEE J Transl Eng Health Med. 2021;9:1–13 https://doi.org/10.1109/JTEHM.2021.3098173.

21. Haugg F, Elgendi M, Menon C. Assessment of blood pressure using only a smartphone and machine learning techniques: a systematic review. Front Cardiovasc Med. 2022;9:894224 https://doi.org/10.3389/fcvm.2022.894224.

22. Zhao H, Zhang X, Xu Y, et al. Predicting the risk of hypertension based on several easy-to-collect risk factors: a machine learning method. Front Public Health. 2021;9:619429 https://doi.org/10.3389/fpubh.2021.619429.

23. Ravaut M, Sadeghi H, Leung KK, et al. Predicting adverse outcomes due to diabetes complications with machine learning using administrative health data. npj Digit Med. 2021;4(1):24 https://doi.org/10.1038/s41746-021-00394-8.

24. Yuan A, Lee AY. Artificial intelligence deployment in diabetic retinopathy: the last step of the translation continuum. Lancet Digit Health. 2022;4(4):e208–e209 https://doi.org/10.1016/S2589-7500(22)00027-9.

25. Xue J, Min F, Ma F. Research on diabetes prediction method based on machine learning. J Phys Conf Ser. 2020;1684:012062 https://doi.org/10.1088/1742-6596/1684/1/012062.

26. Eng D, Chute C, Khandwala N, et al. Automated coronary calcium scoring using deep learning with multicenter external validation. npj Digit Med. 2021;4(1):88 https://doi.org/10.1038/s41746-021-00460-1.

27. van Velzen SGM, Lessmann N, Velthuis BK, et al. Deep learning for automatic calcium scoring in CT: validation using multiple cardiac CT and chest CT protocols. Radiology. 2020;295(1):66–79 https://doi.org/10.1148/radiol.2020191621.

28. Mu D, Bai J, Chen W, et al. Calcium scoring at coronary CT angiography using deep learning. Radiology. 2022;302(2):309–316 https://doi.org/10.1148/radiol.2021211483.

29. Rim TH, Lee CJ, Tham YC, et al. Deep-learning-based cardiovascular risk stratification using coronary artery calcium scores predicted from retinal photographs. Lancet Digit Health. 2021;3(5):e306–e316 https://doi.org/10.1016/S2589-7500(21)00043-1.

30. Siva Kumar S, Al-Kindi S, Tashtish N, et al. Machine learning derived ECG risk score improves cardiovascular risk assessment in conjunction with coronary artery calcium scoring. Front Cardiovasc Med. 2022;9:976769 https://doi.org/10.3389/fcvm.2022.976769.

31. DeFilippis AP, Young R, Carrubba CJ, et al. An analysis of calibration and discrimination among multiple cardiovascular risk scores in a modern multiethnic cohort. Ann Intern Med. 2015;162(4):266–275 https://doi.org/10.7326/M14-1281.

32. Ward A, Sarraju A, Chung S, et al. Machine learning and atherosclerotic cardiovascular disease risk prediction in a multi-ethnic population. Npj Digit Med. 2020;3(1):125 https://doi.org/10.1038/s41746-020-00331-1.

33. Kakadiaris IA, Vrigkas M, Yen AA, Kuznetsova T, Budoff M, Naghavi M. Machine learning outperforms ACC/AHA CVD risk calculator in MESA. J Am Heart Assoc. 2018;7(22):e009476 https://doi.org/10.1161/JAHA.118.009476.

34. Brennan HL, Kirby SD. Barriers of artificial intelligence implementation in the diagnosis of obstructive sleep apnea. J Otolaryngol Head Neck Surg. 2022;51(1):16 https://doi.org/10.1186/s40463-022-00566-w.

35. Tanphiriyakun T, Rojanasthien S, Khumrin P. Bone mineral density response prediction following osteoporosis treatment using machine learning to aid personalized therapy. Sci Rep. 2021;11(1):13811