Artificial intelligence methods and systems for medical consultation

This scoping study explores the role of Human-in-the-Loop Artificial Intelligence (HITL AI) in precision agriculture and evaluates the benefits of using human expertise in combination with Artificial Intelligence (AI) systems in decision-making within modern smart agricultural environments. The development of Artificial Intelligence, Machine Learning, Internet of Things, and robotics has significantly impacted modern agriculture by providing automated crop monitoring, disease detection, yield prediction, and smart farm management systems. However, Artificial Intelligence systems also face challenges in terms of understanding, interpretability, flexibility, and trustworthiness in modern smart agricultural environments. This study is based on the literature regarding human-in-the-loop systems, human-centric Artificial Intelligence systems, and collaborative robotics systems in the context of smart agriculture. The structured scoping study methodology has been followed to identify and evaluate studies regarding Artificial Intelligence systems in smart agricultural environments, with a focus on automation-centric Artificial Intelligence systems and human-centric Artificial Intelligence systems within the context of Agriculture 5.0 concepts. The study concludes that although automation-centric AI systems show high accuracy in simulated smart agricultural environments, Human-In-The-Loop (HITL) AI systems show higher robustness in smart agricultural environments. Explainability in AI has shown significant potential in supporting the effectiveness of HITL AI systems in decision-making within smart agricultural environments. The study also identifies some important gaps in the literature regarding HITL AI systems in smart agriculture. The study concludes that for the development of modern smart precision agriculture, collaborative intelligence within smart agricultural environments is necessary to create sustainable smart agriculture systems.

Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): an international, paired, non-inferiority, confirmatory study

Artificial intelligence (AI) systems are increasingly being used in consumer as well and business applications. Companies are using them, Governments are using them, public security, medical researchers, you and I – everybody seems to be somehow touched. As it becomes more advanced and ubiquitous in our lives, it is critical that AI developers and users understand the ethical implications and challenges posed by this powerful technology. This paper explores the key ethical skills and considerations that are essential for navigating the new world of AI. Artificial Intelligence (AI) presents significant ethical dimensions across various sectors, including research, business, and daily life. It is important to appreciate the major ethical dimensions of AI and how they impact AI systems. Policies and guidelines needs to be laid out for development and usage of such systems. Automated Decisions/Responsibility: No matter how intelligent a AI system might perceived to be, it is always prone to making errors. So, AI system that make automated decisions will always be dangerous. Therefore, it is important to ensure that there is always "human intervention" or "human review" for outcomes/decisions made by AI systems. Determining accountability for AI-generated actions or content remains a significant challenge, emphasizing the need for ethical governance frameworks. There has to be human accountability.Fairness/Bias: AI systems can perpetuate existing biases present in training data, leading to discriminatory outcomes.Transparency: The opaque nature of many AI algorithms complicates accountability, making it difficult to understand decision-making processes.Privacy: AI's capability to process vast datasets can infringe on individual privacy rights, necessitating stringent data protection measuresHuman Dignity: The risk of AI-generated content resembling existing works raises ethical issues regarding originality and authorship.

We discuss how over the last 30 to 50 years, Artificial Intelligence (AI)\nsystems that focused only on data have been handicapped, and how knowledge has\nbeen critical in developing smarter, intelligent, and more effective systems.\nIn fact, the vast progress in AI can be viewed in terms of the three waves of\nAI as identified by DARPA. During the first wave, handcrafted knowledge has\nbeen at the center-piece, while during the second wave, the data-driven\napproaches supplanted knowledge. Now we see a strong role and resurgence of\nknowledge fueling major breakthroughs in the third wave of AI underpinning\nfuture intelligent systems as they attempt human-like decision making, and seek\nto become trusted assistants and companions for humans. We find a wider\navailability of knowledge created from diverse sources, using manual to\nautomated means both by repurposing as well as by extraction. Using knowledge\nwith statistical learning is becoming increasingly indispensable to help make\nAI systems more transparent and auditable. We will draw a parallel with the\nrole of knowledge and experience in human intelligence based on cognitive\nscience, and discuss emerging neuro-symbolic or hybrid AI systems in which\nknowledge is the critical enabler for combining capabilities of the\ndata-intensive statistical AI systems with those of symbolic AI systems,\nresulting in more capable AI systems that support more human-like intelligence.\n

Although the categorization of ultrasound using the Breast Imaging Reporting and Data System (BI-RADS) has become widespread worldwide, the problem of inter-observer variability remains. To maintain uniformity in diagnostic accuracy, we have developed a novel artificial intelligence (AI) system in which AI can distinguish whether a static image obtained using a breast ultrasound represents BI-RADS3 or lower, or BI-RADS4a or higher, to determine the medical management that should be performed on a patient whose breast ultrasound shows abnormalities. To establish and validate the AI system, a training dataset consisting of 4,028 images containing 5,014 lesions and a test dataset consisting of 3,166 images containing 3,656 lesions were collected and annotated. We selected a setting that maximized the area under the curve (AUC) and minimized the difference in sensitivity and specificity by adjusting the internal parameters of the AI system, achieving an AUC, sensitivity, and specificity of 0.95, 90.0%, and 88.5%, respectively. Furthermore, based on 30 images extracted from the test data, the diagnostic accuracy of 20 clinicians and the AI system was compared, and the AI system was found to be significantly superior to the clinicians (McNemar test, p < 0.001). Then, we conducted a trial to introduce the system for use in clinical practice. Physicians reviewed the images and determined whether they were BI-RADS3 or lower, or BI-RADS4a or higher. Next, the classification was performed again for the same images concerning the AI diagnosis. At this time, the initial judgment was allowed to be overturned. We checked whether there was any difference in the diagnostic accuracy, sensitivity, and specificity before and after reviewing to the AI diagnosis. Reviews by 24 physicians were evaluated: 4 Japanese Breast Cancer Society breast specialists, 5 non-specialists and physicians with experience treating more than 40 cases of breast cancer, and 15 non-specialists and physicians with no experience treating more than 40 cases of breast cancer. The average rate of accuracy before confirming the AI diagnosis increased to 73.1% after confirming the AI diagnosis (p=0.00548), compared to 69.3% on average before the AI diagnosis. Compared to practice experience, the accuracy increased from an average of 77.1% to 79.6% for the 9 physicians who were breast specialists or who had treated 40 or more cases of breast cancer. For the 15 physicians with less than 40 breast cancer cases, the average rate of accuracy increased from 64.7% to 69.2%. Furthermore, sensitivity increased significantly to an average of 99.7% after reviewing of the AI diagnosis from an average of 88.8% prior to reviewing the AI-diagnosis.(p< 0.01). Specificity increased from an average of 62.4% to 63.8% (p=0.433) after reviewing AI diagnosis. We showed that our AI system, when applied to clinical practice and used by physicians, contributes to the improvement of diagnostic accuracy. Our results indicated that our AI diagnostic system was sufficiently accurate to be used in the clinical practice. Citation Format: Erina Odani, Tetsu Hayashida, masayuki kikuchi, Aiko Nagayama, tomoko seki, maiko takahashi, Akiko Matsumoto, Takeshi Murata, Rurina Watanuki, Takamichi Yokoe, Ayako Nakashoji, Hinako Maeda, Tatsuya Onishi, Sota Asaga, Takashi Hojo, Hiromitsu Jinno, Keiichi Sotome, Akira Matsui, Akihiko Suto, Shigeru Imoto, Yuko Kitagawa. Establishment of the breast ultrasound support system using deep-learning system [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P1-05-06.

BackgroundThe surgical difficulty of laparoscopic cholecystectomy (LC) for acute cholecystitis (AC) and the risk of bile duct injury (BDI) depend on the degree of fibrosis and scarring caused by inflammation; therefore, understanding these intraoperative findings is crucial to preventing BDI. Scarring makes it particularly difficult to perform safely and increases the BDI risk. This study aimed to develop an artificial intelligence (AI) system to indicate intraoperative findings of scarring in LC for AC.Materials and methodsAn AI system was developed to detect scarred areas using an algorithm for semantic segmentation based on deep learning. The training dataset consisted of 2025 images extracted from LC videos of 21 cases with AC. External evaluation committees (EEC) evaluated the AI system on 20 cases of untrained data from other centers. EECs evaluated the accuracy in identifying the scarred area and the usefulness of the AI system, which were assessed based on annotation and a 5-point Likert-scale questionnaire.ResultsThe average DICE coefficient for scarred areas between AI detection and EEC annotation was 0.612. The EEC’s average detection accuracy on the Likert scale was 3.98 ± 0.76. AI systems were rated as relatively useful for both clinical and educational applications.ConclusionWe developed an AI system to detect scarred areas in LC for AC. Since scarring increases the surgical difficulty, this AI system has the potential to reduce BDI.

To present questions about poisoning to 4 artificial intelligence (AI) systems and 4 clinical toxicologists and determine whether readers can identify the source of the answers. To evaluate and compare text quality and level of knowledge found in the AI and toxicologists' responses. Ten questions about toxicology were presented to the following AI systems: Copilot, Bard, Luzia, and ChatGPT. Four clinical toxicologists were asked to answer the same questions. Twenty-four recruited experts in toxicology were sent a pair of answers (1 from an AI system and one from a toxicologist) for each of the 10 questions. For each answer, the experts had to identify the source, evaluate text quality, and assess level of knowledge reflected. Quantitative variables were described as mean (SD) and qualitative ones as absolute frequency and proportion. A value of P .05 was considered significant in all comparisons. Of the 240 evaluated AI answers, the expert evaluators thought that 21 (8.8%) and 38 (15.8%), respectively, were certainly or probably written by a toxicologist. The experts were unable to guess the source of 13 (5.4%) AI answers. Luzia and ChatGPT were better able to mislead the experts than Bard (P = .036 and P = .041, respectively). Text quality was judged excellent in 38.8% of the AI answers. ChatGPT text quality was rated highest (61.3% excellent) vs Bard (34.4%), Luzia (31.7%), and Copilot (26.3%) (P .001, all comparisons). The average score for the level of knowledge perceived in the AI answers was 7.23 (1.57) out of 10. The highest average score was achieved by ChatGPT at 8.03 (1.26) vs Luzia (7.02 [1,63]), Bard (6.91 [1.64]), and Copilot (6.91 [1.46]) (P .001, all comparisons). Luzia and ChatGPT answers to the toxicology questions were often thought to resemble those of clinical toxicologists. ChatGPT answers were judged to be very well-written and reflect a very high level of knowledge.

Artificial intelligence (AI) systems are now prevalent in our daily lives and hold promise for transforming high-stakes fields such as healthcare. Medical AI systems are showing significant potential to support diagnostics and treatment recommendations. As these systems play an increasingly significant role in clinical decision-making, ensuring transparency in their design, operation, and outcomes is essential for building trust among key stakeholders, including patients, providers, developers, and regulators. However, many systems still function as "black boxes," making it challenging for users-such as clinicians, patients, and other stakeholders-to interpret and verify their inner workings. Here, we examine the current state of transparency in medical AIs, identifying key challenges and risks these opaque systems pose. After motivating the need for transparency in all aspects of the machine learning pipeline, from training data to model development to model deployment, we explore a range of techniques that promote explainability throughout the pipeline while highlighting the importance of continual monitoring and system updates to ensure that AI systems remain reliable over time. Finally, we address the need to overcome barriers that inhibit the integration of transparency tools into clinical settings and review regulatory frameworks that prioritize transparency in emerging AI systems. Through this survey, we aim to increase awareness of current challenges and offer actionable insights for stakeholders, such as researchers, clinicians, and regulators, on how to build trustworthy and ethically responsible AI healthcare solutions.

Artificial Intelligence (AI) systems involve diverse components, such as data, models, users and predicted outcomes. To elucidate these different aspects of AI systems, multifaceted explanations that combine diverse explainable AI (XAI) methods are beneficial. However, popularly adopted user-centric XAI evaluation methods do not measure these explanations across the different components of the system. In this position paper, we advocate for an approach tailored to evaluate XAI methods considering the diverse dimensions of explainability within AI systems using a normalised scale. We argue that the prevalent user-centric evaluation methods fall short of facilitating meaningful comparisons across different types of XAI methodologies. Moreover, we discuss the potential advantages of adopting a standardised approach, which would enable comprehensive evaluations of explainability across systems. By considering various dimensions of explainability, such as data, model, predictions, and target users, a standardised evaluation approach promises to facilitate both inter-system and intra-system comparisons for user-centric AI systems.

Physicians are required to work with rapidly growing amounts of medical data. Approximately 62% of time per patient is devoted to reviewing electronic health records (EHRs), with clinical data review being the most time-consuming portion. To determine whether an artificial intelligence (AI) system developed to organize and display new patient referral records would improve a clinician's ability to extract patient information compared with the current standard of care. In this prognostic study, an AI system was created to organize patient records and improve data retrieval. To evaluate the system on time and accuracy, a nonblinded, prospective study was conducted at a single academic medical center. Recruitment emails were sent to all physicians in the gastroenterology division, and 12 clinicians agreed to participate. Each of the clinicians participating in the study received 2 referral records: 1 AI-optimized patient record and 1 standard (non-AI-optimized) patient record. For each record, clinicians were asked 22 questions requiring them to search the assigned record for clinically relevant information. Clinicians reviewed records from June 1 to August 30, 2020. The time required to answer each question, along with accuracy, was measured for both records, with and without AI optimization. Participants were asked to assess overall satisfaction with the AI system, their preferred review method (AI-optimized vs standard), and other topics to assess clinical utility. Twelve gastroenterology physicians/fellows completed the study. Compared with standard (non-AI-optimized) patient record review, the AI system saved first-time physician users 18% of the time used to answer the clinical questions (10.5 [95% CI, 8.5-12.6] vs 12.8 [95% CI, 9.4-16.2] minutes; P = .02). There was no significant decrease in accuracy when physicians retrieved important patient information (83.7% [95% CI, 79.3%-88.2%] with the AI-optimized vs 86.0% [95% CI, 81.8%-90.2%] without the AI-optimized record; P = .81). Survey responses from physicians were generally positive across all questions. Eleven of 12 physicians (92%) preferred the AI-optimized record review to standard review. Despite a learning curve pointed out by respondents, 11 of 12 physicians believed that the technology would save them time to assess new patient records and were interested in using this technology in their clinic. In this prognostic study, an AI system helped physicians extract relevant patient information in a shorter time while maintaining high accuracy. This finding is particularly germane to the ever-increasing amounts of medical data and increased stressors on clinicians. Increased user familiarity with the AI system, along with further enhancements in the system itself, hold promise to further improve physician data extraction from large quantities of patient health records.

The use of artificial intelligence (AI) systems is significantly increased in the past few years. AI system is expected to provide accurate predictions and it is also crucial that the decisions made by the AI systems are humanly interpretable i.e. anyone must be able to understand and comprehend the results produced by the AI system. AI systems are being implemented even for simple decision support and are easily accessible to the common man on the tip of their fingers. The increase in usage of AI has come with its own limitation, i.e. its interpretability. This work contributes towards the use of explainability methods such as local interpretable model-agnostic explanations (LIME) to interpret the results of various black box models. The conclusion is that, the bidirectional long short-term memory (LSTM) model is superior for sentiment analysis. The operations of a random forest classifier, a black box model, using explainable artificial intelligence (XAI) techniques like LIME is used in this work. The features used by the random forest model for classification are not entirely correct. The use of LIME made this possible. The proposed model can be used to enhance performance, which raises the trustworthiness and legitimacy of AI systems.

Introduction The potential of AI systems to support pre-hospital clinical decision-making is significant. However, whilst such algorithms are increasingly available, far less attention has been paid to understanding the impact of such systems on clinical performance. This study had two aims; first, to compare the performance of expert clinicians against an AI system in a real-world clinical setting and second, to assess the impact of augmenting expert clinical prediction with an AI system. Method We performed a prospective study at two UK Air Ambulances services over a six-month period. Expert pre-hospital clinicians’ judgement of the risk of Trauma Induced Coagulopathy (TIC) in injured patients was assessed and compared to the performance of an AI system. Two TIC risk predictions were generated for every patient: an AI prediction and a human prediction. Results Overall, 51 expert clinicians were enrolled in the study providing 184 patient interactions for analysis. Aim 1 The AI system performed better than clinicians; higher discrimination [AUROC 0.87 (0.79, 0.95) versus 0.83 (0.74, 0.92)] better calibration [0.37 (-0.14, 0.89) versus -1.19 (-1.73, -0.65)] and more accurate [Brier Skill Score 0.34 (0.19, 0.48) versus 0.00 (-0.41, 0.30)]. Aim 2 Risk prediction was better in all performance metrics when clinicians were assisted with the AI system [AUROC 0.88 (0.80, 0.95) versus 0.83 (0.74, 0.92)] Conclusions AI systems can improve human risk prediction in the pre-hospital setting. In settings of low resources where a lack of senior clinical expertise may affect outcomes, the benefit of implementing predictive AI is substantial.

The framework of Input and Output Privacy aids in conceptualization of data privacy protections, providing considerations for situations where multiple parties are collaborating in a compute system (Input Privacy) as well as considerations when releasing data from a compute process (Output Privacy). Similar frameworks for conceptualization of privacy protections at a systems design level are lacking within the Artificial Intelligence space, which can lead to mischaracterizations and incorrect implementations of privacy protections. In this paper, we apply the Input and Output Privacy framework to Artificial Intelligence (AI) systems, establishing parallels between traditional data systems and newer AI systems to help privacy professionals and AI developers and deployers conceptualize and determine the places in those systems where privacy protections have the greatest effect. We discuss why the Input and Output Privacy framework is useful when evaluating privacy protections for AI systems, examine the similarities and differences of Input and Output privacy between traditional data systems and AI systems, and provide considerations on how to protect Input and Output Privacy for systems utilizing AI models. This framework offers developers and deployers of AI systems common ground for conceptualizing where and how privacy protections can be applied in their systems and for minimizing risk of misaligned implementations of privacy protection.

Evaluation of physician performance using a concurrent-read artificial intelligence system to support breast ultrasound interpretation

LAW IN DIGITAL REALITY