Artificial intelligence (AI) is poised to be a world-changing technology. Understanding the basic tenets of AI can help clinicians navigate the anticipated changes it will ultimately bring to health care and the subsequent impact its use will have on patient care. Accounting firm PwC estimates that by the year 2030, adoption of AI will result in an increase of global gross domestic product by 14% ($15.7 trillion) compared to 2017. An estimated 42% of this increase is predicted to be secondary to improvements in economic productivity; however, the remaining 58% (>$9 trillion) is anticipated to arise from increased consumer demand for “higher quality, personalized products and services.”¹ Health care is no exception. As a sector, the Centers for Medicare and Medicaid Services reported health care expenditures comprised 19.7% of total U.S. gross domestic product in 2020, an increase from 17.6% in 2019.²

In the short term, AI is predicted to impact patient care by automating many operations such as processing for medical insurance claims or facilitating automated appointment scheduling.¹ Tangible improvements in efficiency such as these will likely increase patient demand for physician services. With growth in both the overall population and adults >65 years of age,³ as well as the simultaneous contraction in the urology workforce (expected to contract from 3.99 urologists per 100,000 individuals in 2020 to a projected century low of 3.3 per 100,000 between 2030-2035⁴), the increases in productivity enabled by AI-based solutions are going to be critical for maintaining equitable, affordable, and timely access to urological care in the United States.

The field of AI is thought to have been founded in 1956 by the late Dr John McCarthy, believed by many to be the father of AI.⁵ Although varying definitions exist, in his own words, AI is described as “the science and engineering of making intelligent machines, especially intelligent computer programs.”⁶ Machine learning (ML), a subfield of AI science, focuses on utilizing data to detect, extrapolate patterns, and make predictions. Clinical medicine is abundant when it comes to patient data. Many AI projects explored in the medical field to date have utilized a variety of ML algorithms. One such example is illustrated with ML and predicting sepsis. Concerns regarding the high sensitivity of systemic inflammatory response syndrome criteria or poor sensitivity of quick Sequential Organ Failure Assessment and the subsequent delayed diagnosis have led many authors to investigate the utility of ML in this setting.⁷ Chen et al, in a meta-analysis comprising 24 retrospective series, found that the majority of ML-based models had high accuracy for early identification of sepsis.⁷ Implementation of such predictive algorithms into the clinical workflow can enable clinicians to identify true sepsis early on and avoid false-positive predictions.

Deep learning (DL), a subfield of ML, utilizes vast quantities of data and computational power to facilitate many permutations of interactions between input data before providing a final output value.⁸ This computational process facilitates all variables to interact (“neural network”) and generates models capable of representing the complexity of real-world input data. DL models are commonly utilized in medical image interpretation and analysis models.

An example of the utility of DL in clinical practice can be seen with prostate cancer. In 2022, there were an estimated 270,000 new cases diagnosed in the United States.⁹ The predominant method of diagnosis, prostate biopsy, results in generation of at least 10-12 individual tissue samples that require pathological review. Since the Gleason score is the primary modality used by urologists in determining treatment course, an accurate score is vital. However, Gleason scoring is also subject to both intra- and interobserver variability. Bulten et al demonstrated the utility of DL-based image analysis (eg pathology slide) by training a model similar in performance to pathologists.¹⁰ The authors reported their model outperformed 10 out of 15 pathologists in 1 experiment and had high concordance with readings from 2 independent pathologists.

Like pathology, radiology also involves an abundance of imaging data that require expert interpretation. In assigning PI-RADS scores, proper segmentation of a lesion is essential; however, the task remains a challenge due to patient and/or imaging-related factors.¹¹ Using a DL model, Bardis et al demonstrated in a series of 242 patients undergoing prostate MRI that AI-based segmentation of both the transitional and peripheral zones can be accurately performed in an automated fashion, thereby enhancing radiologist workflow. Image-based decision support may also benefit urologists directly. Shkolyar et al reported a highly precise DL model for identification of bladder tumors in real time under white light cystoscopy.¹² Augmentation of common urological procedures such as cystoscopy with AI can enhance clinical decision-making and may ultimately translate to improved patient care.

Every minute, patient data, as a manifestation of the human condition, grow in data centers around the world. The digitization of disease is perhaps the key to unlocking the many mysteries in medicine that remain unsolved. As physicians, we must be at the forefront of facilitating the growth of AI in medicine as a form of augmentation rather than replacement. Patient care will always require a human touch, but as generations of patients and physicians age, AI and its timeless nature may facilitate continued learning through the generations that remains elusive today.

Since 2002, the AUA Residents and Fellows Committee has represented the voice of trainee members. The Committee’s mission is to address the educational and professional needs of urology residents and fellows and promote engagement with the AUA. The Committee welcomes your input and feedback! To contact us, or inquire about ways to be involved, please email rescommittee@AUAnet.org.