Digital Transformation in Personalized Medicine with Artificial Intelligence and the Internet of Medical Things

Introduction

In a broad sense, artificial intelligence (AI) aims to simulate human intelligence using a machine (Hammond, 1973; Waltz, 1982). As early as 1950, Mr. Turing “proposed to consider the question ‘can machine think?’” (Turing, 1950). Then, the concept of AI was first proposed at the Dartmouth Summer Research Project on Artificial Intelligence at Dartmouth College in Hanover in 1956 with the statement that “Every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it” (McCarthy et al., 2006).

Traditionally, AI can be divided into artificial narrow intelligence (ANI) and general artificial intelligence (GAI). ANI (or weak AI or narrow AI) aims to solve simple tasks such as imaging analysis or disease diagnosis using collected static data for training using machine learning (ML) techniques for structured data or using natural language processing (NLP) methods to extract and convert information from unstructured data such as clinical record into structured data for ML (Ghahramani, 2015). Algorithms used in ML include, for example:

Logistic regression,

GAI is more advanced and general form of AI, which aims to match or surpasses human intelligence, that is, the ability to “think” or learn on its own and to evolve and become better tomorrow than today. However, it is still in its early stages and yet to see real-world implementation.

With the advances of computing power and wireless capabilities, the Internet of Medical Things (IoMT), the application of the Internet of Things (IoT) in medicine, is dramatically changing the way medicine is practiced and delivered. With the ability to generate, collect, analyze, or transmit health data or images efficiently, IoMT greatly improves the efficiency of the health care (Dimitrov, 2016; Gershenfeld et al., 2004).

In this review, we will first focus on the current applications of ANI and IoMT in personalized medicine in oncology and COVID-19 as two examples, and then move on to describe the use of real-time data for the development of digital personalized medicine.

AI and Personalized Oncology

The IBM Watson health (https://www.ibm.com/watson-health) has pioneered the application of AI in medicine using both ML and NLP techniques and achieved good performance. For example, in oncology, Aikemu et al. (2020) compared the performance of the Watson for Oncology (WFO) from IBM and a multidisciplinary team (MDT) using concordance rate between the two for treatment recommendations for colorectal cancer patients. They found that the concordance rates for treatment recommendations were 97%, 93%, 89%, 87%, and 100% for neoadjuvant, surgery, adjuvant, first-line, and second-line treatment groups, respectively, suggesting robust performance of the WFO (Aikemu et al., 2020).

In another study for lung cancer, Kim et al. (2020) compared the performance of the WFO from IBM and a MDT for four treatment recommendations (surgery, radiotherapy, chemoradiotherapy, and palliative care) for lung cancer using Cohen's kappa value. They found that the concordance between MDT and WFO was 92.4% (k = 0.881, p < 0.001) of all cases. However, the concordance differed according to clinical stages with 100% concordance (k = 1.000) for stage IV non-small cell lung carcinoma (NSCLC) and extensive disease SCLC, but at lower concordance for stage III NSCLC (80.8%, k = 0.622) and relatively low in stage II NSCLC (83.3%, k = 0.556) and limited disease SCLC (84.6%, k = 0.435).

A critical challenge in improving the performance of AI and adoption of AI in routine hospital setting beyond research or even clinical research setting is the understanding and processing of the unstructured patients' data. The experience and problem encountered during early adoption of the IBM Watson system at the University of Texas M. D. Anderson Cancer Center in Houston highlighted the shortcoming of the AI in digesting written case reports, doctor's notes, and other nonstructured text-based information generated in the hospital setting (Schmidt, 2017). As we move further into the IoMT for generating more structured digital records, AI such as the IBM Watson system will better perform the tasks in real-life clinical setting.

Bibi et al. (2020) simulated IoMT technology by uploading blood smear images to a cloud computer for AI diagnosis of leukemia using two CNN approaches: dense CNN (DCNN) and residual CNN (RCNN). They showed that AI-based subtype classification of leukemia supersedes all the previous approaches. Based on the size and shape of WBC, leukemia can be classified into acute and chronic, and then each into subtypes lymphoid and myeloid leukemia, which resulted in four subtypes: acute lymphocytic leukemia (ALL), chronic lymphocytic leukemia (CLL), acute myeloid leukemia (AML), and chronic myeloid leukemia (CML). Bibi et al. (2020) showed that, for ALL, both CNN approaches achieved 100% accuracy, and for AML, CLL, and CML, the diagnosis accuracies were 99.65% and 99.91%, 99.73% and 99.91%, and 99.73% and 100% for DCNN and RCNN methods, respectively.

McKinney et al. (2020) developed an AI system to improve correct interpretation of mammograms and found that the AI system dramatically improved the accuracy of reading with an improved the area under the receiver operating characteristic curve (AUC-ROC) of 11.5% compared with that by an average radiologist in both United Kingdom and the United States, where the study was conducted. Furthermore, the AI system reduced the necessity of the second reading by 88%, thus showing great potential in improving the accuracy and efficiency of breast cancer screening (McKinney et al., 2020).

Personalized risk prediction improved diagnosis efficiency and offered personalized management of breast cancer. Kakileti et al. (2020) developed a personalized risk score named Thermalytix risk score (TRS) using AI to analyze thermal images to help identifying a high-risk target population for regular screening for early-stage breast cancer detection. The TRS achieved an AUC of 0.89, which is much better than the AUC of 0.68 by an age normalized risk score, and the TRS correlates with the likelihood of malignancy, and further categorized the individuals into four risk groups, providing a personalized guidance on screening frequency needed (Kakileti et al., 2020).

Personalized treatments in oncology would be the best choice and AI is playing critical roles as the biology of cancer is complicated. For example, for radiation therapy of breast cancer, multiple factors must be taking into consideration, including breast anatomical volumes, for personalized radiation dosing while minimizing damage to other organs and the locoregional recurrence risks for risk-adjusted dosing.

Furthermore, high- and low-risk areas for recurrence are influenced by multiple factors, including individual factors (e.g., age and comorbidity), disease-specific factors (e.g., initial tumor location within the breast, type of surgery, stage, or grade), treatment-related factors (e.g., extent of surgery, systemic treatments, and radiation dosing), and progress monitoring factors (e.g., changes in biomarker concentrations, changes in tumor immunity profiles, and changes in tumor sizes). As the influencing factors are multiple, AI including DL would be of great help to build an intelligent system for personalized and optimized treatments.

AI and Personalized Treatment for COVID-19

The SARS-COV-2 pandemic posed one of the greatest challenges to the mankind and strained the health care systems worldwide. COVID-19 (coronavirus disease 2019), a systemic disease caused by SARS-COV-2 virus, not only affects the respiratory system but also other organ systems from brain to blood vessels (Wadman et al., 2020), and, if not treated timely, could deteriorate quickly and leading to death in some cases. Timely diagnosis of the infection is critical in the management of the disease, from containing the spread of the virus to correct treatment of the patients. Digital health technologies, including AI and the IoMT, offer great help for COVID-19 diagnosis and management to improve patient care, and for combating the pandemic.

For example, Yousefi developed a reagent-free SARS-COV-2 sensing and diagnosis with 5 min readout time using only a sensor-modified electrode chip in unprocessed patient saliva (Yousefi et al., 2021). In addition, ML and AI were also employed to help the diagnosis of the COVID-19. Murugan et al. applied CNN models using X-ray images for differential diagnosis among COVID-19, pneumonia, and normal (Murugan and Goel, 2021) and achieved 94.07% accuracy, a 98.15% sensitivity, and a 91.48% specificity (Murugan and Goel, 2021).

Timely prediction of the COVID-19 disease severity is critical in offering personalized treatment to help those needed intensive care and mechanical ventilation while preventing the waste of resources for milder patients. Based on the COVID-19 tracking project (https://covidtracking.com/) as of March 5, 2021, only 3.1% COVID-19 patients required hospitalization and only 20% have required care in the intensive care unit (ICU); 6.8% required mechanical ventilation among the hospitalized patients.

Patel et al. (2021) developed a Random Forest classifier using sociodemographic data, clinical data, and blood panel profile data collected at the time of admission to predict the need for ICU with AUC = 0.80 and mechanical ventilation with AUC = 0.82. Patel's model relied only on sociodemographic data, data acquired from a physical examination, and laboratory marker data obtained from a blood draw that were available in resource-poor clinical settings or community-level hospital where radio imaging might not be readily accessible. They further simplified the model using only the top five features [C-reactive protein (CRP), d-Dimer, procalcitonin, SpO2, and respiratory rate] for predicting ICU need with similar performance of AUC of 0.79 (0.72, 0.85) for predicting ICU and AUC of 0.83 for mechanical ventilation (Patel et al., 2021).

Similarly, Heldt et al. (2021) also built an XGBoost model with AUC of 0.0.84–0.87 for predicting the need for ICU and mechanical ventilation using patients' initial presentation at the emergency department. They further demonstrated that the patient's age and oxygenation status were most predictive of poor outcome, and blood lactate and deoxyhemoglobin levels also are key contributors to the model (Heldt et al., 2021). Many other studies have found similar performance in using ML to risk stratification of the COVID-19 patients (Fernandes et al., 2021; Schoning et al., 2021) with initial physical assessment and blood test data or with additional information from the chest computed tomography scans (Paiva Proenca Lobo Lopes et al., 2021; Pourhomayoun and Shakibi, 2021; Quiroz et al., 2021).

Rapid testing and early identification of COVID-19 is critical in containing the spread of SARS-COV-2 virus. Mendels et al. (2021) developed a xRCovid smartphone application using CNN algorithm to help reading, and correctly interpreting the positive and negative calls for rapid COVID-19 diagnostic tests using blood samples. They showed that across 11 different types of COVID-19 rapid tests and 3344 tested samples, only 18 false negatives and 5 false positives were observed, achieving an overall sensitivity of 98.9% and specificity of 99.7%. In this case, the application of AI and IoMT greatly simplified the test interpretation process and help harvesting the benefits of the COVID-19 rapid tests in combating the pandemic.

Toward Real-Time Personalized Medicine Using AI

In the past, AI was trained with previously collected and annotated data, and then used to make prediction using the new data. However, with the development of the real-time imaging capabilities such as real-time colonoscopy, on the spot training and deployment of AI offers real-time decision from real-time data. For example, endoscopy health care workers were called upon to make the decision to remove the polyps (adenoma and benign neoplastic lesions) during colonoscopy. This is being enabled by convergence of advances in several fronts such as the availability of biosensors, wireless connectivity, AI, and the IoT (Lin and Wu, 2020; Ozdemir, 2020) (Fig. 1).

OPEN IN VIEWER

FIG. 1.

Digital transformation in personalized medicine and its technical pillars, including the IoT, AI, sensors, and wireless connectivity (modified from Özdemir, 2020). AI, artificial intelligence; IoT, Internet of Things.

Now, with the help of AI techniques such as deep neural networks or CNN, real-time intraprocedural automated detection of polyps based on optical diagnosis during live colonoscopy is possible (Chao et al., 2019; Mori et al., 2021; Wittenberg and Raithel, 2020). The AI techniques improved the detection and the differentiation of neoplastic and non-neoplastic lesions in the gastrointestinal (GI) track. Several randomized controlled trials have indicated that AI tools possibly increased the adenoma detection rate by roughly 50% and contributed to a 7–20% reduction of colonoscopy-related costs (Mori et al., 2021). Several commercial platforms for computer-aided detection and computer-aided diagnosis have obtained regulatory approval. These include, for example, the GI Genius from Medtronic Corp., Ireland; EndoBRAIN-EYE Cybernet Corp., Japan; CAD EYE, Fujifilm Corp., Japan; Discovery, Pentax Corp., Japan.

Cancer harbors mutations that could be identified by the next-generation sequencing (NGS) technologies. Drug resistance often evolved during treatments with changes in genomic landscapes and expression profiles. Mody et al. conducted integrative clinical exome (tumor and germline DNA) and transcriptome (tumor RNA) sequencing for 28 patients with hematological malignancies and 63 patients with solid tumors. They found that an average of 46% of the patients (54% with hematological malignancies and 43% with solid tumors) had actionable findings that changed their cancer management (Mody et al., 2015).

In the end, personalized actions, including changes in treatments or risk counseling, were initiated in 23 of the 91 (25%) patients based on actionable integrative clinical sequencing findings (Mody et al., 2015). NGS allows profiling of cancer genomes for identifying ongoing genetic aberrations that could be used to develop tailored and personalized treatments based on the actionable mutations.

Conclusions

Digital transformation is impacting every facet of science and society, not least because there is a growing need for digital services and products with the COVID-19 pandemic (Lin and Wu, 2020). But the need for digital transformation in diagnostics and personalized medicine field cuts deeper. In the past, personalized/precision medicine initiatives have been unable to capture the patients' experiences and clinical outcomes in real-time and in real-world settings. The availability of wearable smart sensors, wireless connectivity, AI, and the IoT is changing the personalized/precision medicine research and implementation landscape. Digital transformation in poised to accelerate personalized/precision medicine and systems science in multiple fronts, which signals, in our view, the rise of digital personalized medicine research and practices in systems science.