New Horizons: Artificial Intelligence Tools for Managing Osteoporosis

Abstract

Osteoporosis is a disease characterized by low bone mass and microarchitectural deterioration leading to increased bone fragility and fracture risk. Typically, osteoporotic fractures occur at the spine, hip, distal forearm, and proximal humerus, but other skeletal sites may be affected as well. One of the major challenges in the management of osteoporosis lies in the fact that although the operational diagnosis is based on bone mineral density (BMD) as measured by dual x-ray absorptiometry, the majority of fractures occur at nonosteoporotic BMD values. Furthermore, osteoporosis often remains undiagnosed regardless of the low severity of the underlying trauma. Also, there is only weak consensus among the major guidelines worldwide, when to treat, whom to treat, and which drug to use. Against this background, increasing efforts have been undertaken in the past few years by artificial intelligence (AI) developers to support and improve the management of this disease. The performance of many of these newly developed AI algorithms have been shown to be at least comparable to that of physician experts, or even superior. However, even if study results appear promising at a first glance, they should always be interpreted with caution. Use of inadequate reference standards or selection of variables that are of little or no value in clinical practice are limitations not infrequently found. Consequently, there is a clear need for high-quality clinical research in this field of AI. This could, eg, be achieved by establishing an internationally consented “best practice framework” that considers all relevant stakeholders.

Osteoporosis is defined as a systemic skeletal disease characterized by low bone mass and microarchitectural deterioration of bone tissue with a consequent increase in bone fragility and susceptibility to fracture (1). According to criteria recommended by the World Health Organization (WHO), the operational diagnosis of osteoporosis is based on bone mineral density (BMD) as assessed by dual x-ray absorptiometry (DXA) of the hip, the spine, or the so-called one-third radius (2, 3). Osteoporosis can thus be diagnosed if an individual's BMD is equal to or less than −2.5 standard deviations below the mean normal BMD value of healthy young adults (ie, at a T-score ≤ −2.5). However, one of the main limitations of this approach lies in the fact that the majority of fractures occur at T-scores of −1.0 to −2.5 (ie, at osteopenic BMD) or even above −1.0 (ie, at normal BMD), compromising the sensitivity of this “gold-standard” method and its potential role as a screening tool (4). Fractures at the spine (vertebrae), hip (proximal femur), shoulder (proximal humerus), and wrist (distal forearm, distal radius) have been shown to be associated with an increased risk of subsequent fracture, reduced quality of life, disability, and, except for distal forearm fracture, increased mortality (5-8). Accordingly, they have also been referred to as major osteoporotic fractures. The average lifetime risk of a 50-year-old woman to suffer a major osteoporotic fracture has been estimated at close to 50%, and in men 22%, and, worldwide, osteoporosis causes some 9 million fractures annually, resulting in an osteoporosis fracture every 3 seconds (9, 10).

The term artificial intelligence (AI) in its current sense was most likely coined in the mid-1950s, when a group of mathematicians and cognitive and computer scientists convened at a conference at Dartmouth College (U.S.). While the conference itself fell short of the participants’ expectations, it can nevertheless be considered the initial spark for the AI research boom that followed. This boom, however, was interrupted at least twice by periods of research slump, sometimes also referred to as “winters of AI research,” the first of which lasted from the mid- to the late 1970s and the second from the late 1980s to the early 1990s. Both “winters” were preceded by barely encouraging research results, which in turn led to a decrease in funding of AI-related research projects. Fortunately, with an almost exponential growth of computational power, research and funding began to pick up again after that. In 1997, an IBM computer called IBM Deep Blue® beat the world chess champion Gary Kasparov, and in 2011 another IBM computer, called Watson®, beat 2 of the all-time most successful human players of the game Jeopardy in front of millions of television viewers. Undoubtedly, these and many subsequent highlights in the development of AI have formed the perfect foundation for AI-based research efforts in human medicine. In fact, substantial progress across many areas of human medicine has been made in the past decade (11, 12). In general, AI in medicine can be divided into 2 subtypes, namely virtual and physical, with the former including eg, imaging solutions and treatment decision support tools, and the latter, eg, intelligent prosthesis and robot-assisted surgery (13). In regard to the management of osteoporosis, the virtual subtype of AI currently plays the main role, with solutions available (or in development) to facilitate diagnosis, fracture risk assessment, fracture detection, bone quality assessment, and treatment decision (14) (Fig. 1).

Figure 1.

Selection of currently available artificial intelligence solutions for osteoporosis management.

Some Basics of AI in Clinical Medicine

Very much simplified and given that currently there is no internationally consented definition available, AI constitutes a system that combines computational power with data sets (ideally big data) to enable problem-solving. A typical branch of AI is machine learning, which uses various algorithms to learn from data, thus clearly differing from the (human) attempt to explicitly write a specific computer program to achieve a specific task (15). Machine learning (ML) can be based on different learning approaches, the most important of which are supervised learning and unsupervised learning (16). In the past decade, supervised ML turned out to be most efficient and has thus become the main pillar of AI-supported healthcare applications. To train a system that is based on supervised learning, the machine has to be fed with data that are already available and robust, because the quality of these input data will determine the quality of the output. For example, to train a machine to detect hip fractures on a conventional radiograph, one must feed the ML algorithm with a set of conventional hip radiographs that contain fractured and nonfractured hips. In addition, one has to tell the system which hips are fractured and which are not, a process that is also referred to as “annotation” of images. In general, the more such annotated images the algorithm is fed, the better it will become at detecting fractures. Deep learning is a more powerful subset of ML with an architecture that is similar to the human brain in that multiple layers of “neurons” are interconnected with each other, forming a so-called neural network (17). Among the NNs currently available, the so-called convolutional neural networks (CNNs) are those most widely used in healthcare applications.

In general, the development of a specific AI algorithm requires a training data set and a test data set. In some cases, a third data set is put aside for validation purposes only. Ideally, the entire available data set, such as a set of radiographs, is split randomly into these 2 or 3 sets prior to the development process. This ensures that data from the same patient is not used for more than 1 of these data sets, a constellation that is often termed “data leakage” and that would compromise test and validation results and lead to misclassification and misdiagnosis in clinical practice (18). The training data set is used for algorithm training and usually involves sets of characterizing data points, also referred to as “features,” and corresponding predictions, respectively (16). Such features can be simply pixels or voxels in radiographs or clinical diagnoses or laboratory parameters as extracted from electronic medical records. Typically, the training data set is much bigger than the test data set, with a ratio, eg, of 80:20 up to 60:40. However, currently there are no international standards available, eg, in regard to a minimum sample size of such data sets. In order to find the best model, the developed algorithm has to be validated by a cross-validation process (19). The classic form of cross-validation is termed holdout method, and it simply involves the data training set and the data test set. However, this method is not effective for, eg, comparing multiple models. For this and other purposes, a so-called k-fold cross-validation is used. To put it simply, in this very method the entire data set is divided into k (ie, a specified number of) groups of data. In its simplest form, this could again be 1 training data set and 1 test data set. However, in contrast to the holdout method, both the test data set and the training data set are used interchangeably. In other words, if for example k was 10, the entire data set will be divided into 10 groups, and 10 separate models will be built. In the first iteration, nine-tenth of the data set will be used for training purposes, while one-tenth of the data will be kept aside for testing. In the second iteration, another group will be put aside for testing, while the remaining 9 groups will serve for training purposes etc. This process will then be repeated 8 more times.

Once a newly developed model is found to be sufficient, eg, for hip fracture detection, its performance is tested/validated against human performance. In the case of a hip fracture detection algorithm, this could be expert physicians such as radiologists and/or orthopedics. Performance results are usually expressed using classic performance metrics, such as sensitivity, specificity, and area under the curve (AUC) of the receiver operating characteristic. If more than 1 model is tested, the model with the best performance might then be chosen for further testing and validating in clinical studies, approval by regulatory authorities, and implementation into the respective clinical application.

Opportunistic Diagnosis of Osteoporosis

The reasons to apply opportunistic methods for the operational diagnosis of osteoporosis are severalfold. For example, the gold-standard method for diagnosis of osteoporosis according to the recommendations of the WHO is based on measurement of BMD using DXA (3). However, access to DXA devices is largely dependent upon their availability. Aside from the number of DXA units available within a country and some patient-related obstacles such as traveling time and transportation opportunities, regulatory constraints and barriers to reimbursement play an important role as well. In a recent survey it was shown that the number of DXA units per million of a country's general population ranges from less than 10 (which is considered inadequate) up to more than 50 (20). In most countries worldwide, the number of available DXA devices is considered inadequate. Thus, there is a clear need for applying alternative methods that ideally would be more widely available, such as conventional radiography.

For example, in dental medicine, and particular dental implantology, panoramic radiographs have been used for many years to gain information about quantitative and qualitative properties of the mandibular bone, albeit with rather unsatisfactory results (21, 22). In addition, it has been recommended to refer patients to medical professionals if osteoporosis is suspected from panoramic radiographs (23). Thus, there is a good rationale for the development of AI-supported tools to improve diagnostic performance of such opportunistic osteoporosis screening methods. For example, in a preliminary study, with no subsequent full study published so far, a set of different CNNs was applied to panoramic radiographs to diagnose osteoporosis (23). Depending on the CNN used, AUC values ranged from 0.98 to 0.99. However, instead of using DXA as the gold-standard reference technology, diagnosis of osteoporosis was made if cortical erosions were observed in the same radiographs by 2 experienced oral and maxillofacial radiologists. In another study, 4 different CNN models were investigated, and it was found that transfer learning and fine tuning of such CNNs can improve diagnostic performance considerably, up to an AUC of 0.86 (24). However, not only do the results of this study sound less promising, but, in addition, the clinical relevance is severely compromised by the fact that no information was provided about the reference technology used.

AI-based software solutions for opportunistic diagnosis of osteoporosis have also been developed using conventional radiographs from the hand and wrist. In fact, in 1 study cortical radiogrammetry from the third metacarpal bone shaft and cancellous texture analysis from distal radius were used to train and test the algorithm, and DXA was used as a reference standard (25, 26). The developed software exhibited an accuracy of close to 89%, which could make it an interesting option for wider clinical use.

Other methods that have been proposed for opportunistic diagnosis of osteoporosis are, eg, computed tomography (CT), quantitative computed tomography (QCT), and quantitative ultrasound. In general, such methods provide T-scores that would allow a categorization into “normal”, osteopenic,” and “osteoporotic.” However, since all of these alternative technologies capture different bone properties resulting in different gradients of risk, T-scores obtained cannot simply be used interchangeably with T-scores derived from DXA (27). On the other hand, it should be kept in mind that, irrespective of the gold-standard status of DXA, there is evidence that BMD assessed by opportunistic QCT of the spine may show a higher association with the risk of incident vertebral fractures than T-scores measured by DXA (28).

Considering such aspects, a number of studies have strived to improve the sensitivity, specificity, and accuracy of such opportunistic approaches by applying AI-supported software tools. For example, in 1 study chest CTs made for lung cancer screening were used to measure BMD of the spine and assign patients to 1 of the WHO categories by using a fully automated AI algorithm (29). The diagnostic performance was found to be very good, with an AUC of 0.83 for osteopenia and 0.97 for osteoporosis. Furthermore, it was found that with every 10 HU increase of CT values, the risk of osteopenia decreased by 32% to 44% and the risk of osteoporosis by 61% to 80%. The authors concluded that routine chest CTs in combination with AI are of great value in opportunistic screening for osteopenia and osteoporosis. However, the limited clinical relevance of these findings lies in the fact that DXA-based BMD measurement of the spine is performed using lumbar vertebrae L1-L4, while chest CT-based BMD measurement is mainly based on thoracic vertebrae. Considering this methodological weakness, in 1 study a CNN model was developed to predict BMD from abdominal CT scans using DXA of the lumbar spine as reference standard (30). Osteoporosis was correctly diagnosed with an AUC of 0.965 for the internal validation data sets and 0.970 for external data sets. Similar promising results were found in another abdominal/pelvic CT study that used DXA as a reference standard as well (31). In another study designed to develop an AI-based tool for diagnosis of osteoporosis using abdominal or pelvic CTs, DXA was not used as reference standard, limiting the clinical relevance of the otherwise promising study results (32).

The numerous imaging-based AI tools for opportunistic diagnosis of osteoporosis have led to a first systematic review and meta-analysis recently (33). A total of 7 studies including more than 3000 patients were found eligible for inclusion. Using a random effects model, the pooled sensitivity was 0.96 and the pooled specificity was 0.95. However, as pointed out correctly by the authors, results should be interpreted with caution due to the high risk of bias in patient selection and high heterogeneity. This being said, it should also be noted that out of the 7 studies included, only 3 had used DXA technology as a reference standard.

Lastly, bone marrow fat fraction as assessed by magnetic resonance imaging has been shown to be associated with abnormal bone density (34). However, the clinical utility of this approach is limited by the fact that image segmentation has to be performed manually. In a study involving some 200 healthy volunteers, a fully automated end-to-end radiomics pipeline using image segmentation via CNN was developed (35). Using QCT of the lumbar spine as the reference method, the developed bone marrow fat fraction map radiomics achieved excellent performance in predicting osteopenia and osteoporosis.

Detection of Osteoporotic Fractures

One of the mainstays in osteoporosis management is fracture detection in conventional radiographs. Therefore, it is not surprising that a considerable number of AI-supported fracture detection software tools has been developed so far. Typically, most of them involve 1 specific CNN algorithm, such as Inception, Xception, or DenseNet, but in some cases a combination of different CNNs, sometimes also referred to as a an “ensemble,” is used (36). In general, it has been shown that such AI tools are reliable in fracture diagnosis and that they have a high diagnostic accuracy, similar to that of expert physicians such as radiologists or orthopedics (37). Such evidence from individual studies is also supported by the findings of a couple of systematic reviews and meta-analyses (29, 38). However, in 1 meta-analysis it was found that the diagnostic performance (ie, the pooled sensitivity and specificity) was less convincing when all studies eligible for inclusion were considered, as opposed to the results obtained from a subgroup analysis that only included the “long-bone” group without vertebrae, clavicle, and ribs. This finding is of clinical relevance as it provides evidence that AI-supported fracture detection tools may work less reliably if the skeletal site of interest together with its surrounding tissue is of more complex structure, making not only the correct classification of the region of interest more difficult but also detection of fracture lines themselves. In this regard, many such classification problems have been shown to be due to insufficient (sizes of) training data sets used (39). In the more recent meta-analysis, aside from heterogeneity and study bias issues, significant flaws in study methods were pointed out (38). For example, among all studies included, only 1 study provided a sample size calculation. Furthermore, only 1 study had provided clinicians with background clinical information, so that the clinician performance was most likely underestimated and the whole study process not representative of a real-world setting.

Vertebral fractures are the most abundant osteoporotic fractures, but only one-third of them would come to clinical attention immediately, with the rest being detected more or less by chance or in the course of diagnostic evaluation of chronic back pain (40). In conventional radiographs of the spine or the chest, particularly if made for other reasons than exclusion or diagnosis of vertebral fracture, detection rate of vertebral fractures has been shown to be low (41, 42). Consequently, awareness programs such as the Capture the Fracture initiative have been developed worldwide to improve expert physicians’ diagnostic performance in this regard (43). Aside from increasing awareness, technical support in the form of AI-based software tools appears to be a logical add-on. Therefore, automated detection of vertebral fractures in conventional chest and spine radiographs has recently come into focus of AI developers. As an example, an AI-based software program was developed for vertebral fracture detection on elderly women's lateral chest radiographs (44). The software considers a semiquantitative categorization of these fractures according to the Genant classification that includes 3 different grades of height-loss, namely mild (20-25%; Grade 1), moderate (25-40%; Grade 2), and severe (>40%; Grade 3) (45). The severity of vertebral fractures, ie, the amount of height loss, has important clinical implementation in that the risk of subsequent fracture is highest in patients with a Grade 3 vertebral fracture. Furthermore, it has been shown that antifracture efficacy of osteoporosis drugs depends on the severity (and number) of vertebral fractures (46). Overall, diagnostic performance (ie, sensitivity, specificity, and accuracy) of this tool appears promising, although performance results for vertebral fractures with only Grade 1and Grade 2 losses were less convincing.

Since conventional radiographs of the spine taken immediately after a trauma sometimes do not show morphological changes, advanced medical imaging methods such as magnetic resonance tomography (MRT) and CT are used in addition. For example, MRT is frequently applied to detect bone marrow edema as an indicator of the recency of a vertebral fracture. Clinically, such information can serve as a decision support if vertebral augmentation is considered (47). In this regard, an AI-based algorithm was developed for automated detection of fresh osteoporotic vertebral fractures recently, whereby “fresh” was defined as a period of 3 months after the respective injury (48). The finally chosen AI algorithm was a combination of 4 different CNN models (ie, an “ensemble”), which yielded the best performance and was comparable to that of 2 experienced spine surgeons. The image output of this AI tool provides a color-coded classification of vertebrae into “normal,” “fresh fractured,” and “old fractured” (Fig. 2). The authors conclude that the algorithm developed in this study may contribute to the daily care of osteoporosis patients by helping to reduce misdiagnosis of fresh osteoporotic vertebral fractures, particularly in hospitals without radiologists or spine surgeons. However, independently of some methodological limitations such as a relatively small training data set, 1 of the main limitations is that the algorithm was not trained to identify pathological fractures such as those associated with metastatic bone disease.

Figure 2.

(A) Image output after automated fracture detection and classification into “normal” (white), “fresh fractured” (red), and “old fractured (blue)”. (B) Base; The original image. http://creativecommons.org/licenses/by/4.0.

First AI-based algorithms for automatic detection of vertebral fractures in CT scans were published almost a decade ago (49). While these and most of the algorithms developed thereafter technically required multiple segmentation of each vertebra, a more recently developed algorithm involving a specific CNN model allows extraction of radiological features from each slice in a full CT scan (50). Extracted features are then processed through a so-called feature aggregation module to make the final diagnosis for the full CT scan. The algorithm, which can be applied to CT scans of the chest, the abdomen, and the pelvis, yields a diagnostic accuracy of close to 90%, with an overall performance comparable to that of radiologists.

However, the most recent and probably most advanced approach to detect vertebral fractures from abdominal and chest CT scans involves a 3-dimensional voxel classification method, which requires segmentation neither of the individual vertebra nor of the full CT scan (51). The 3D method developed here shows excellent performance with an AUC of 95% for patient-level fracture detection and an AUC of 93% for vertebra-level fracture detection.

Fracture Risk and Fracture Prediction

For estimation of an individual's 10-year fracture probability, the most widely used tool worldwide is FRAX®, which is an online-available, free-of-charge fracture risk assessment tool encompassing more than 80 country- and region-specific versions (52). In fact, this tool covers about 80% of the world population, and it is recommended for use by almost all national osteoporosis guidelines (52-54). It is primarily based on clinical risk factors such as prevalent fracture, parental hip fracture, glucocorticoid use, smoking, and alcohol abuse, and it can be used with or without DXA-based BMD results (55). In the past few years, a number of AI-based models have been developed for fracture risk prediction. For example, in a recently prospective community-based cohort study, a novel AI-based fracture prediction model was developed, and its performance was compared to that of the country-specific version of FRAX® (56). Out of 3 different models developed, the 1 that performed best exhibited an AUC of 0.688 for fracture prediction, which was significantly better than that achieved by FRAX®. The top predicting risk factors were BMD of the total hip, the lumbar spine, and the femoral neck. Surprisingly, even factors such as a subjective arthralgia score, serum creatinine, and homocysteine were listed higher than conventional predictors such as age or prevalent fracture. In another study that was based on longitudinal data from a larger cohort, CNN-based models were developed using conventional spine radiographs (57). It was found that the model that only used baseline radiographs provided vertebral fracture risk prediction comparable to that of FRAX®. Using the data sets that included DXA results, the predictive performance of the AI model was even higher than that of FRAX®. However, again such results have to be interpreted with caution, because FRAX® provides 10-year fracture probability, while none of the studies mentioned here was designed for predictions covering a period of 10 years. In addition, it should be kept in mind, that irrespective of the impressive performance of these AI models in the population studied, the results cannot be simply extrapolated to other populations, as baseline fracture risks may differ markedly among different populations (58).

Bone Properties Beyond BMD

Bone quality is determined not only by BMD, which in fact provides information primarily about bone quantity and its degree of mineralization, but also by its geometry, microarchitecture, and tissue composition (59). At the macroscale, bone strength can be assessed (in vitro or postmortem) by whole-bone mechanical testing, in which bone is loaded to failure under compression, bending, or torsion (60). In recent years, noninvasive estimation of bone strength has also become possible in vivo (as opposed to the invasive in vivo method of microindentation) by using finite element models that can be integrated into the software of different imaging modalities such as QCT (61). In a recent AI study, material properties and geometric features of vertebral body were extracted from QCT images obtained from Asian male subjects, and an ML algorithm was developed aiming to propose a convenient and practical method for clinical strength prediction of vertebral bodies (62). Study results were promising in terms of prediction ability and consistency, and the authors concluded that the algorithm developed here may have great potential for noninvasive assessment of vertebral fracture risk. However, again it should be kept in mind that these study results cannot be generalized to different ages, ethnicities, or female sex. In another recent study that involved DXA results and clinical variables from a female population with prevalent fractures, a model was developed to identify patients prone to subsequent fragility fractures (63, 64). The developed CNN based Bone Strain Index reached a predictive accuracy of close to 80% with a sensitivity of 75% and a specificity of 84%. The authors concluded that the Bone Strain Index appears to be a useful DXA index in identifying those patients who are at risk of further vertebral fractures. In a study based on ultrasound technology (ie, ultrasound attenuation) a CNN-based model was developed to estimate micro-architectural properties of cortical bone (65). The final model had the ability to predict (ie, quantify) the micro-architectural parameter of cortical porosity with high accuracy. This might be useful particularly in treatment monitoring to detect, eg, bone anabolic properties of a bone active drug without exposure to radiation. Another interesting approach was aimed at developing a ML algorithm for direct automated quantitative measurement of the spine that is used for spine assessment in several clinical entities such as osteoporosis, intervertebral disc degeneration, or disc herniation (66). The model was tested against T1-and T2-weighted MR images and achieved an impressive performance with mean absolute errors of 1.22 ± 1.04 mm and 1.24 ± 1.07 mm for the selected 30 MRT-based lumbar spinal indices. Undoubtedly, the model developed in this study has potential to support clinical decision making, but, similar to most of the early AI-based solutions, it will need further confirmation/validation in large-scale studies.

The DXA-based Trabecular Bone Score (TBS) is another indirect measure of bone strength. It has been shown to capture morphometric properties of vertebral bone by using pixel based gray-level variations, thus providing information beyond that provided by DXA BMD alone (67). For this reason, and irrespective of some limitations, TBS was integrated into the fracture risk prediction tool FRAX® to adjust the calculated 10-year fracture probability if indicated (68). No larger scale prospective clinical studies involving AI-supported TBS have been performed so far, but in a smaller retrospective study involving patients treated with the osteoanabolic drug teriparatide, an artificial neural network analysis revealed a significant amelioration of TBS, which might explain the known reduction in fracture risk independently of BMD (69).

Treatment Decision Support

The question of whom to treat when and which osteoporosis drug to use to reduce a patient's fracture probability most efficiently has been a matter of debate in the past 3 decades. Currently, there is worldwide consensus among most osteoporosis-related scientific societies that patients who suffered a low-trauma fracture should receive osteoporosis treatment with proven antifracture efficiency (70-72). This approach is based on strong evidence that a first fracture—sometimes also referred to as sentinel fracture—is a strong predictor of subsequent fracture, with the highest probability within the first 12 to 24 months after fracture (7). Individuals who suffered an osteoporotic fracture in the past 12 to 24 months are therefore considered as being at very high (or imminent) risk (73). Furthermore, bone anabolic drugs such as teriparatide or romosozumab are currently recommended as first-line treatment in these patients because of their early and more pronounced antifracture efficacy compared to antiresorptive drugs (7, 74). There is, however, less consensus on when to initiate treatment in a patient without prevalent fracture, although current concepts are based on the logical inference that if a patient with a prevalent fracture must receive osteoporosis treatment, anyone else must be treated if his or her fracture probability is at least equal to that of a patient of the same age and sex (72). However, different tools are available to calculate a patient's fracture risk, and categorization of risk, eg, into high or very-high, may differ depending on the respective guidelines and/or recommendations (70-72). This being said, a considerable proportion of osteoporosis patients who receive “adequate” treatment with proven antifracture efficacy fail to respond (75, 76).

Considering these facts, it is not surprising that in the recent past software developers have made increasing efforts to make use of big data available in the form of electronic medical records as provided by different healthcare systems and providers. For example, a recently developed AI algorithm to predict treatment-related BMD response was based on electronic medical records of more than 15 000 osteoporosis patients followed over a period of 10 years (77). In addition to 5200 International Classification of Diseases codes, the algorithm considered about 30 000 BMD results and more than 3500 different drugs, but notably only 7 different laboratory parameters, with total alkaline phosphatase the only one of some relevance in regard to bone turnover. Neither vitamin D nor an established bone resorption or formation marker was included. Nevertheless, out of 7 different ML algorithms developed, the one with the best performance to predict treatment response in terms of BMD increase showed an receiver operating characteristic of 0.70 and an accuracy of 0.69. Aside from any relevant clinical information, a typical printout of this AI-based software also provides a list of potentially eligible drugs, including information on which of these drugs would most likely be associated with the highest BMD increase (Fig. 3).

Figure 3.

Typical report as provided for an osteoporosis patient with inadequate treatment response. Drug names were replaced by consecutive numbers. Drug #3 is recommended because it suggests the highest likelihood of adequate bone mineral density response. http://creativecommons.org/licenses/by/4.0.

In summary, a large number of AI algorithms have been developed in the past few years to facilitate the management of osteoporosis including diagnosis, fracture risk assessment, fracture detection, bone quality assessment, and treatment decision. Where applicable, performance comparison between such solutions and human physician experts show similar results or are even in favor of the AI algorithm. Specifically, radiomics including AI-supported opportunistic methods for BMD assessment together with clinical data appear to have a great potential for early detection of patients who are at increased risk of fracture. However, even if study results appear promising at a first glance, methodological approaches behind a newly developed AI algorithm should always be carefully read and critically appraised. Use of inadequate reference standards or selection of features (ie, variables) that are of little or no value in clinical practice are limitations not infrequently found in AI development studies. Also, it is not always evident that the AI algorithm chosen or developed would result in a performance superior to that of simple traditional statistical methods. Consequently, there is a clear need for high-quality clinical research in the field of AI in osteoporosis management. This could be achieved by, eg, establishing an internationally consented best practice framework that considers AI developers, osteoporosis experts, and their respective scientific societies as well as healthcare authorities including those involved in approval processes. Once these quality requirements are met, the potential of AI to revolutionize osteoporosis management may be fully unlocked, and benefits to physicians, patients, and healthcare in general are likely to become better visible and more meaningful in the future.

Funding

This work was supported by the Medical University of Graz, Austria.

Data Availability

Data sharing is not applicable to this article as no data sets were generated or analyzed during the current study.