Artificial intelligence in paediatric endocrinology: conflict or cooperation

AI in paediatric diabetes

AI holds significant potential in the management of paediatric type 1 diabetes. A recent trial aimed at evaluating the efficacy and safety of frequent insulin dose adjustments, guided by an automated AI-based Decision Support System (AI-DSS), compared to adjustments guided by physicians from specialist academic diabetes centres in glucose level control. The study demonstrated that the AI-DSS system was comparable to physicians in terms of insulin dose adjustment every three weeks and the duration spent within the targeted glucose range was non-inferior across the AI-DSS groups in comparison with the physician-guided patients. This potentially provides an opportunity to facilitate automated insulin calculations improving paediatric diabetes workflow and management [7].

One of the challenges for children and young people with diabetes is the early recognition of hypoglycaemic or hyperglycaemic episodes facilitating early intervention. Paediatric patients suffering from type 1 diabetes are at greater risk for developing acute rather than chronic complications, compared to adult patients [8]. Given the risk of long-term complications from poor glycaemic control, it is pivotal that paediatric patients achieve optimal HbA1c levels and reduce glycaemic variability. In an ongoing trial aiming to enrol 64 patients less than 18 years old who are already wearing a continuous glucose monitor, participants also wore an additional non-invasive wearable device for recording ECG data (Medtronic Zephyr BioPatch). The objective of the study was to validate an established AI model to automatically detect hypoglycaemic events by using a few ECG heartbeats recorded with the wearable device. The proposed system automatically learnt patterns in the ECG, discriminating between heartbeats recorded during low or normal glucose levels in the same subject [9].

AI has also been utilized to detect diabetic retinopathy (DR) in paediatric patients [10]. DR, a leading contributor to global blindness and impaired vision, affects approximately 100 million individuals with retinopathy threatening vision in around 30 million [11, 12]. These numbers are projected to rise in the future, thus early detection through screening in young people is recommended to facilitate timely intervention. Screening for DR is recommended in children with type 1 diabetes from 11 years of age and with a diabetes duration of 3–5 years. Children with type 2 diabetes patients are advised to undergo screening soon after diagnosis, and annually thereafter. In 2018, the FDA approved the marketing of an AI device to detect diabetic retinopathy catalysing automated DR detection [13]. Employing a non-mydriatic fundus camera with autonomous AI has proven to be a secure and efficient method for diabetic eye examinations in the young people, achieving a sensitivity of 85.7 % and a specificity of 79.3 % [14].

Automated algorithms for DR detection offer numerous advantages over human-based screening. They don’t experience fatigue and can analyse thousands of images daily, often delivering results within seconds to minutes after image capture. Moreover, scaling automated DR screening programs primarily involves acquiring additional hardware, making it a highly feasible process. However, systems are currently costly and, therefore, utilisation is limited for low-income countries where DR incidence is high, and human graders are still required to judge atypical or low-quality images, and to provide quality assurance for automated systems, particularly with the low specificity of some automated AI DR systems [15], albeit sensitivity and specificity has improved significantly [16]. Other AI applications that will improve future diabetes care and outcomes include the use of AI to predict hypoglycaemic and hyperglycaemic events [17, 18], predicting those who are at risk of gestational diabetes, predicting gestational diabetes onset and determining the most effective course of treatment, in-turn reducing the risk of maternal and infant complications, such as pre-eclampsia, birth trauma, large-for-gestational age infants and neonatal hypoglycaemia [19, 20], and the use of AI in closed loop systems utilising continuous glucose monitoring data to predict insulin dose [21, 22].

AI and imaging in paediatric endocrinology

Computer vision and deep learning techniques have been employed to process the large data sets held within radiological images to improve diagnostic accuracy, efficiency and patient care. As with the study to diagnose lung cancer in primary care, an AI model reading radiographs for diagnosis needs to be shown to be superior or at least more efficient with equivalent capability to the skills of experienced radiologists before its use can be recommended. This use of AI is encapsulated in a new field of radiology called radiomics. Radiomics involves the extraction of numerous quantitative features from medical images including intensity, shape, size, texture and spatial relationships within the region of interest and the use of AI to analyse these data. Before feature extraction, medical images are pre-processed to ensure consistency and quality. This may involve noise reduction, normalisation, registration and segmentation to define the region of interest accurately.

AI algorithms can automatically segment and annotate medical images, making it easier for radiologists to identify and analyse regions of interest and detect abnormalities. AI can also support image reconstruction and enhancement, thus improving image quality and reconstructing high-resolution images from low-resolution scans, enhancing diagnostic accuracy and aiding clearer visualization of structures. AI models can federate and analyse medical image data alongside clinical data to predict patient outcomes, disease progression and treatment responses aiding in personalised treatment planning and decision-making [23]. This now extends to a new field of medicine called radiogenomics by which AI algorithms can correlate imaging data with clinical and genomic data diagnosis and treatment planning [24], [25], [26], [27].

At present, the use of AI in radiology to improve diagnostic accuracy and predict disease progression is limited and each AI model needs to be critically appraised in appropriate clinical settings beyond trialling training and test sets. Deep-learning AI models have been developed to differentiate malignant tumours from benign thyroid nodules in adults and children. However, presently these AI models demonstrate high sensitivity but low specificity [28, 29]. AI models have also been applied to accurately predict bone age with a high level of diagnostics accuracy across models ranging from 85 to 95 % across an age range of 1.5–18 years within 1 year of the bone age calculated by expert radiologists [30–33], with the most accurate methods predicting bone age within 4.3–4.5 months of pre-defined bone ages assessed by radiologists [34]. In relation to the prediction of final adult height in children entering puberty early, AI algorithms using the TW3 (Tanner Whitehouse 3) methodology for bone age assessment appear to work best in predicting final height in females with early puberty [35]. At present, however, AI algorithms in studies are essentially determining bone age within models that have a value assigned by expert radiologists, therefore, introducing a risk of inaccuracy and bias. Therefore, presently more rapid and efficient calculation of bone age by AI models provides the more compelling argument for using automated bone age assessment. Future work using significantly larger and more comprehensive databases to develop a training dataset of healthy children could be used to develop AI models that independently predict bone age – essentially removing ‘human ground truth’ [36] and eliminating the inherent biases of using data derived from radiologist assessment to provide training sets [37, 38].

AI in puberty and growth

ML approaches have been developed to diagnose central precocious puberty (CPP) in females. The first publication in this field used two ML models to analyse retrospectively collected data on 1757 females who had presented with potential CPP. A total of 966 females had a confirmed diagnosis of CPP based upon a peak LH level ≥10 IU/L or peak LH level ≥5 IU/L combined with a ratio of peak LH to FSH of ≥0.6 and together with the onset of secondary sexual characteristics at the age less than 8 years. The ML algorithms using 19 variables were able to predict CPP puberty with a sensitivity ranging from 77.91 to 77.94 %, specificity ranging from 84.32 to 87.66 % and with the area under the curve (AUC) ranging from 0.88 to 0.90. A smaller sample (n=436) was subsequently tested using a 25 variable ML model including imaging data performed with equivalent accuracy but diminished in accuracy when only using 19 variables, demonstrating the balance between the size of the patient cohort and the number of variables in the models. In both models, the most important predictive variable was LH level, followed by IGF-I and FSH levels. In conclusion, the authors suggested that ML models may be used as a reliable screening tool to positively identify CPP before considering a GnRH stimulation test [39].

Similar models have been created using 15 variables in 4 separate ML models in 161 females with AUC values greater than 0.88 across models [40]. Another approach in predicting CPP has been to use nine variables that were extracted from the participants’ clinical records in a machine learning algorithm including age (years), body weight (Z-score), height (Z-score), BMI (Z-score), obesity, breast development (Tanner stages 1–5), pubic hair (Tanner stages 1–5), onset of menarche and bone age. The remaining five biochemical variables were basal oestradiol, LH, and LH and FSH results from a shortened GnRH stimulation test that had previously been validated for use in diagnosing CPP [41], [42], [43]. The ML model could accurately predict CPP in female participants with a positive predictive value of 0.987, AUC of 0.972 and a specificity of 0.893, demonstrating the potential to reduce sampling standard GnRH testing [44]. Alternative approaches using ML multi-omics approach to investigate the alterations and functional characteristics of gut microbes and blood metabolites in patients with CPP have yielded interesting results. Alterations in the gut microbiome may yield alternative molecular therapeutic markers to diagnose CPP and monitor response to therapy [45].

Facial recognition software is also being applied in medicine to identify subtle facial changes and dysmorphic features that may be present in certain conditions, some of which may be too subtle to be identified by clinicians [46, 47]. In adult patients, ML algorithms have been developed to diagnose acromegaly with an accuracy of around 95 % [47, 48]. Similar findings have been achieved in patients with Cushing syndrome (CS) on retrospective facial analysis [46], although the accuracy of ML models in predicting CS diminishes significantly when a matched sample of patients with high BMI without CS are included in [49]. Facial recognition machine learning models have also been used to identify patients with Turner syndrome, the partial or complete loss of one X chromosome in females. Based upon 68 facial characteristics, one ML model achieved a diagnostic accuracy of 83.4 % on the test set [50]. Using a larger data set, an alternative model using deep convolutional neural networks was able to achieve a higher diagnostic accuracy of 97 % demonstrating the potential value in the use of facial recognition ML in diagnosing Turner syndrome [51]. A similar technique may be applied to diagnosis of other genetic conditions relevant in paediatric endocrinology such as Noonan syndrome [52] and other non-genetic disorders that affect growth such as foetal alcohol spectrum [53]. More broadly, facial recognition programmes have been developed with the potential to differentiate hundreds of genetic conditions such as Deep Gestalt with the ability to differentiate subtypes of conditions within the Noonan syndrome [54] and automated 3-dimensional facial imaging [55].

Interestingly, facial recognition software has also been applied to diagnosing congenital adrenal hyperplasia (CAH), a condition which is not classically characterised by morphological changes in facial features [56]. In this study involving 102 individuals with CAH and 144 individuals in the control group, the utilisation of deep learning techniques resulted in an AUC of 92 % for the accurate prediction of CAH using facial images. Further investigation into facial regions highlighted the significance of the nose and upper face in making the most significant contributions to this distinction. The authors speculated that prenatal organizational or postnatal effects of androgen excess on 11 facial morphologic features in patients with CAH may result in these subtle facial differences [5]. One of the future challenges of using ML facial recognition is the application of these technologies to ethnic subsets and to different age ranges, particularly in younger children given data sets are currently limited and models are highly dependent on comprehensive datasets to refine current systems [46, 57].

The assessment of growth and prediction of future growth are pillars of paediatric endocrinology and AI programmes are showing promise in assisting in this field. The early detection of growth disorders assists with earlier referral to and assessment by paediatric endocrinologists, facilitating timely diagnosis and intervention and potentially better longer-term outcomes. In Finland, around 98 % of the child population participates in a growth monitoring program, with 20 height measurements taken from the first post-birth measurement to aged 12 years. An automated algorithm built into patient electronic health records to identify problems with growth demonstrated an earlier diagnosis of disorders such as coeliac disease and subsequent referral to specialist care in comparison to the previous standardised system [58]. Others have developed an approach using the input of anthropometric data entered into an app by parents with subsequent advice to identify growth disorders and obesity using an embedded AI algorithm. The app facilitated earlier referral to specialists and earlier diagnosis of growth disorders and obesity intervention [59].

An alternative application is the use of ML to predict final adult height (AH) during childhood. To achieve this, comprehensive growth data from longitudinal paediatric cohorts are used to train ML algorithms, with an accuracy of nearly 90 % in predicting AH, but with prediction errors that result in over- or under-estimates for short and tall subjects, respectively. Importantly, the ML approach taken for AH prediction was shown to be non-inferior to traditional methods of calculation [60].

AI in predicting and diagnosing disease

Machine learning is now being applied to predicting and diagnosing diseases in children, which includes paediatric endocrinology. ML models incorporate relevant features (predictors) selected or engineered from the data to represent essential aspects of a child’s presentation, such as genetic markers, vital signs, symptoms and signs. Within the models, feature engineering may involve creating new variables or combining existing ones to enhance predictive accuracy. Subsequently, variables are used in various machine learning algorithms, such as decision trees, random forests, support vector machines, neural networks or deep learning models, to determine the most appropriate model(s) to provide the most accurate prediction based upon the variables. These models learn patterns and relationships between features and target outcomes (e.g. disease presence, severity) during the training process. Trained models are then evaluated using validation datasets to assess that their performance and necessary adjustments are made to optimize predictive accuracy, sensitivity and specificity. Once the models are sufficiently trained and validated, they can predict the likelihood of a child developing a particular disease based on input data. These predictions can aid in early diagnosis, allowing for timely intervention and improved health outcomes.

In an era in which obesity prevalence is high, ML models have been used to predict the onset of pre-diabetes in young people by using a non-invasive method, in an attempt to support individuals with lifestyle modification to prevent the onset of type 2 diabetes in adulthood [61]. ML models have been applied to clinical and biochemical parameters to predict metabolic-associated fatty liver disease [62] (MAFLD – previously known as non-alcoholic fatty liver disease of NAFLD), to identify children and young people that may be at greater risk of developing hepatic complications and supporting targeted interventions [63]. Other ML models have been applied to metabolic profiles to predict future outcomes. Using random forest machine learning models, metabolomic profile at 3 months can predict body composition and adiposity at 2 years of age with a predictive value of 75.8 %, sensitivity of 100 % and specificity of 50 %. These metabolites included several classes of lyso-phospholipids and are influenced by infant feeding type [64, 65].

Another clinical condition in which ML may add value is in the diagnostic process is the differentiation between central diabetes insipidus (CDI) and primary polydipsia, attempting to circumvent the need for the water deprivation and hypertonic saline tests both of which are intensive for young patients. The initial ML model included 56 patient characteristics clinical, biochemical and radiological covariates of which five covariates were identified as crucial for differentiating CDI from primary polydipsia – urine osmolality, plasma sodium and glucose, known transsphenoidal surgery, and known anterior pituitary dysfunction. Although the model was highly accurate, as expected, adding in MRI data improved the model’s accuracy, and pituitary stalk enlargement was found to be the main MRI covariate for the diagnosis of CDI along with the other baseline parameters. The authors of this paper proposed that the ML model may be used to differentiate patients who are more likely to have CDI and consider MRI scanning specifically for this patient subgroup. This is particularly important in young patients where a general anaesthetic is required to conduct the scan. The authors subsequently concluded that their ML algorithm would act as an important clinical decision support tool to help overcome the challenge of differentiating CDI with primary polydipsia [66].

Others have applied an ML approach to the diagnosis on classical adrenal congenital adrenal hyperplasia using the combination of adrenal steroid hormone analysis using liquid chromatography-tandem mass spectrometry (LC-MS/MS), bone age and clinical parameters. Based on selected variables, the ML model was 100 % accurate in differentiating premature adrenarche from late-onset CAH patients. The most significant variables in the ML model were 21-deoxycorticosterone, 17-hydroxyprogesterone and 21-deoxycortisol steroids [67]. Similar findings were seen using an ML model to differentiate adult women with non-classical CAH and polycystic ovarian syndrome [68], demonstrating the value in both models of combining clinical features with a single blood test to identify patients with non-classical CAH, without the need for synacthen (co-syntropin) testing. Other patient groups stand to benefit from the use of AI and ML to support their care and prevent long-term complications. ML algorithms as part of continuous glucose monitoring (CGM) offer significant potential in predicting hypoglycaemic episodes in neonates, infants and children with hyperinsuinaemic hypoglycaemia to prevent irreversible neurodevelopmental complications due to the late identification of hypoglycaemia in this patient group using current methods [69], and the wider applicability of CGM in preterm babies who are prone to adverse neurodevelopmental outcomes from hypo- and hyper glycaemic episodes [70]. However, CGM ML models are currently not accurate enough to apply to real-world settings [71].

ML models have also been applied to determine the factors that predict response of patients with Turner syndrome to recombinant growth hormone (GH), with factors such as chronological age at the start of treatment predicting the height SDS in the first year of therapy, and the increment of height SDS in the first year being an important predictor of height SDS gain after 3 years, providing clinicians with a platform to predict growth hormone response in this patient group [72].

Using AI in clinical decision support

Precision medicine, also known as personalized medicine or individualized medicine, is an innovative approach to healthcare that customizes medical treatment and preventive strategies based on an individual’s unique genetic, molecular, clinical and lifestyle characteristics. The goal of precision medicine is to tailor interventions to each person’s specific needs, optimizing outcomes and minimizing adverse effects. AI enables the categorization of patients into various levels based on disease subtypes, risk factors, demographics and socio-economic attributes. This allows for personalized interventions, tailored to each patient. Integration of factors such as genomics, lifestyle, medical history, therapy response and adherence is now feasible based upon widespread data acquisition from multiple sources [73, 74]. In the future, AI will be used to provide the means to advise clinicians about clinical interventions for paediatric endocrine patients based upon their data and provide personalised and targeted plans. Anthropometric data and data related to growth hormone adherence are already being transmitted via digital platforms through a connected growth hormone delivery device to inform decision-making about future therapeutic changes to improve growth through therapy adherence [75, 76], and a framework has since been conceptualised highlighting the potential for future digital interventions to support the management of children with growth disorders requiring growth hormone therapy [77]. Analyses of adherence data using ML to identify types of patients with reduced adherence have indicated worse adherence in those who infrequently transmit data, don’t change growth hormone device comfort settings and start treatment at an older age in the following 3, 6 or 9 months. This information can be used to provide support to at-risk patients to achieve optimal adherence and, subsequently, improve clinical outcomes with growth hormone therapy [78]. In paediatric type 1 diabetes clinics, AI models have been shown as effective in determining future insulin dose based upon patient data, highlighting the potential for use of these algorithms in and outside the clinic to support therapeutic decision-making [21, 79]. More broadly, clinical decision support systems are being developed to aid clinicians with effective diagnoses and interventions based upon medical records, clinical, laboratory and imaging data [80, 81]. Ultimately, the synergy between AI and clinical decision support has the potential to significantly enhance healthcare delivery for children and young people, empower clinicians, improve efficiency and workflows and ultimately improve patient outcomes by providing evidence-based, personalized and timely clinical decision support.

Challenges with using AI in paediatric populations

Despite its potential, the integration of AI in paediatric endocrinology presents several challenges. Data privacy and protecting children’s health data is crucial. Robust data encryption and compliance with privacy regulations are necessary. Efforts are required to mitigate algorithmic bias and ensure equitable healthcare outcomes, especially in paediatric population and AI algorithms must undergo rigorous clinical validation to demonstrate their safety, efficacy and reliability in the context of age-specific populations. To address this, a recently published framework entitled ACCEPT-AI sets out recommendations for the safe inclusion of paediatric data in AI and machine learning research that address the fundamental parameters of age, consent, assent, communication, equity and protection of data. ACCEPT-AI has been designed to guide researchers, clinicians, regulators and policymakers [82]. The aim of the ACCEPT-AI guidelines is to address the potential algorithmic bias that may occur when using paediatric data. In the realm of machine learning, algorithms depend on diverse or multiple datasets, often referred to as training data. These datasets contain predefined outputs for various individuals or objects. Utilizing this training data, the algorithm constructs a model, which can subsequently be employed to make predictions for other individuals or objects, determining what the correct outputs should be for them. AI systems learn to make decisions based on training data, which can include biased human decisions or reflect historical or social inequities in data sets due to the under- or over-representation of specific groups [83, 84]. Some algorithms run the risk of replicating and even amplifying human biases, particularly those affecting protected groups, leading to erroneous outputs that can result in misclassification and potential harm [85, 86]. Algorithmic bias in paediatric data may result from failure to document age, lack of representation of children in population data sets and inference about paediatric data extrapolated from adult data sets. Given the physiological, anatomical and developmental differences across the paediatric age range and between paediatric and adult groups, failure to train the technology on these differences has the potential to create algorithmic outputs that are not valid, applicable, effective or generalisable across age subgroups. The ACCEPT-AI framework defines age-specific measures to prevent algorithmic bias and attain equitable, ethical and appropriate AI/ML outputs on paediatric data sets [82]. Bajwa et al. have suggested an iterative approach to test AI outputs by evaluating and validating the predictions made by the AI tool to test how well it is functioning based upon three dimensions: statistical validity of the model, clinical utility to demonstrated clinical effectiveness and generalisability and economic utility to test whether there is plausible argument to invest in a novel AI system [87].

To this end, in 2019, the McKinsey Company made recommendations in the development of AI and risk mitigation by taking an approach that involved experts in firstly developing the AI algorithm based upon the conceptual challenge, secondly supporting implementation and delivery of algorithms in the desired setting or pathway and thirdly determining the outputs in relation to social and ethical implications [88]. Ethical review boards now and in the future will be required to assess AI development in line with standards to minimise the downstream effects of biased training data sets [89].

The interface between AI and clinical skills

At a pragmatic level, the question has been raised as to whether AI will replace doctors and, in this context, paediatric endocrinologists. AI should in fact be considered as a complimentary tool rather than a replacement for healthcare professionals with the ability to consistently improve accuracy and precision, efficiency and predictive power, particularly as clinical skills and expertise may vary considerably based on training, experience and the provision of facilities for clinical practice [90]. From the examples provided in this paper, AI amplifies and augments, rather than replaces, human intelligence and the integration of AI technology into healthcare should not be about ‘them vs. us’ in the context of machines [91, 92]. In their present form, AI systems and algorithms are unable to reason, draw upon experience or intuition and do not use emotion or empathy to support decision-making and future planning of patient care. ML has been described as a ‘signal translator’ in which the translation rule is learned directly from the data [92]. However, contrary to the belief that computers can’t emulate emotion, recent research has demonstrated that AI chatbots were able to provide a more empathetic and detailed response to online patient questions compared to physicians [93].

However, given that AI systems are dependent on training data, if a patient with a specific problem presents to a paediatric endocrinology clinic but the clinical history reveals another issue related to either physical or mental health, an AI platform may not be able to understand the issues presented as there is no training data to inform the model. Furthermore, more subtle changes relating to an initial endocrine diagnosis or the progress of a clinical condition may not be identified in an AI model if the training set has not incorporated data on these subtle changes. AI algorithms are also founded on prior knowledge to create the model architecture and to determine how the model learns. In the context of healthcare, problems are often complex, multifaceted and not always easy to define, making rule-based models difficult to manufacture. Disease-based models in paediatric endocrinology are often influenced by confounding socio-economic factors that mean that already small datasets become over-segmented and AI models yield biased results [94].

Importantly in the context of paediatrics and child health, children are not simply biological models, but they have individual needs and vulnerabilities and are also considered in the context of family relationships and dynamics. In this context, the core value of clinical care needs to remain at the heart of the relationship between a patient and a healthcare professional [95]. AI cannot create the doctor–patient relationship or test the strength or weakness of that relationship. The quality of a doctor–patient relationship is based on human interaction, which seeks to create a feeling of commitment and trust in the patient towards their treating physician. Such a commitment depends on the empathy and clinical skills of the physician and cannot yet be created by digital innovation. This provides the opportunity for the physician to make eye contact with the child and carer and to gain a valuable impression of the clinical problem and inter-family relationships. The skill of history-taking includes asking direct questions, listening to the replies and following up on the information received. The nuances of this dialogue are processed and acted on. In this way, a symptom of subtle relevance such as a mild alteration of bowel or respiratory function can be followed up by further direct questions, which could easily be missed by a pre-constructed digital model. Furthermore, the human empathetic input is in shorter supply by virtue of time constraints, which clinicians need to work to combat. Interestingly, a recent perspectives paper by Eric Topol considered the value of AI in the context of ‘restoring the essential humaneness in medicine, primarily by providing the gift of time’, in a way that creates opportunities for clinicians to focus on compassion and empathy. By creating efficiency through automating tasks such as using large language models to answer questions, utilisation of AI to improve diagnostic speed and accuracy and the use of AI algorithms to deliver future care and facilitate monitoring at home, this provides much needed time to focus on the value of patient–caregiver–doctor relationships.

Furthermore, AI could also be used to coach doctors to improve empathy, sensitivity and open-ended questioning. Topol, however, talks of the one-dimensional nature of AI, learning only from text, and lacking the ability to replicate the nuances of social interaction with information derived from all of the human senses [96]. AI has yet to be applied to assessment of leadership in terms of clinical practice. A patient with a complex disorder such as congenital adrenal hyperplasia makes the best progress when seen at every out-patient appointment by the same experienced physician. If the clinic is badly organised, due to lack of physician leadership, and the patient sees a different doctor at each visit, unfortunately not an uncommon situation, progress will be impaired. The ensuing demoralisation of the patient and family, easily translated into poor adherence to a drug regimen, is currently beyond the scope of AI to gauge and correct. The atmosphere within a hospital department is largely determined by the quality of relationships between the leader and his/her staff.

AI can recognise patterns, but it cannot gauge emotions or instincts or psychological variants or human emotions such as doubt, fear, suspicion or confidence. Further still, in the process of clinical examination, an AI model cannot emulate the value of touch in evaluating a child’s presenting signs. During the process of decision-making for interventions and treatments, this often involves a nuanced dialogue between the paediatrician, children and their families to determine the best and personalised approach.

AI models will characteristically rank decisions based upon algorithmic determinants, which may be based on factors such as length of treatment, frequency of medication or side effect profile. Moreover, given the potential inherent bias that AI models may generate, trust in digital tools may be rapidly undermined [97]. For future AI models to emulate a relationship with patient and families, the models must be founded upon the principles of privacy, confidentiality and fairness, and in some way allow informed choices [98]. However, the current ‘paternalistic’ decision-making approach that AI models adopt may not align directly with the choice of the patient, which is often driven by external and environmental factors which influence choice [99].

AI is dependent on data quality and access, which becomes highly relevant in the field of paediatric endocrinology in which children and young people characteristically present with rare conditions, and compiling comprehensive datasets relies on national and often international collaboration. To integrate AI platforms, centres also need the appropriate technical infrastructure, organisational capacity and governance. In respect to governance, the legal frameworks around clinical responsibility are challenging given that AI decision-making can often be perceived as a ‘black box’ process where liability is difficult to apportion on digital technologies [89]. It is this ‘black box’ concept where the inner logic of AI algorithms remains concealed even from their developers and, therefore, creates lack of transparency and trust by patients and families.

Table 1 summarises the benefits and challenges of using AI in paediatric endocrinology.