Fracture Risk Assessment and Clinical Decision Making for Patients with Metastatic Bone Disease

Abstract

Metastatic breast, prostate, lung and other cancers often affect bone, causing pain, increasing fracture risk, and decreasing function. Management of metastatic bone disease (MBD) is clinically challenging when there is potential but uncertain risk of pathological fracture. Management of MBD has become a major focus within orthopedic oncology with respect to fracture and impending fracture care. If impending skeletal related events (SREs), particularly pathologic fracture, could be predicted, increasing evidence suggests that prophylactic surgical treatment improves patient outcomes. However, current fracture risk assessment and radiographic metrics do not have high accuracy and have not been combined with relevant patient survival tools. This review first explores the prevalence, incidence, and morbidity of MBD and associated SREs for different cancer types. Strengths and limitations of current fracture risk scoring systems for spinal stability and long bone fracture are highlighted. More recent CT-based structural rigidity analysis (CTRA) and finite element (FE) analysis methods offer advantages of increased specificity (true negative rate), but are limited in availability. Other fracture prediction approaches including parametric response mapping (PRM) and positron emission tomography/computed tomography (PET/CT) measures show early promise. Substantial new information to inform clinical decision-making includes measures of survival, clinical benefits, and economic analysis of prophylactic treatment compared to after-fracture stabilization. Areas of future research include use of Big Data and machine learning to predict SREs, greater access and refinement of CTRA/FE approaches, combination of clinical survival prediction tools with radiographically based fracture risk assessment, and net benefit analysis for fracture risk assessment and prophylactic treatment.

INTRODUCTION

Management of patients with metastatic bone disease (MBD) for which there is risk of pathological fracture remains a clinical challenge. Preventing pathological fracture through use of prophylactic treatment (PT) would seem to be a reasonable approach for many patients, but determining whether a fracture might occur in the future is often difficult. This is a very active area of orthopedic research and is a topic of increased focus within the orthopedic oncology community that requires contributions from clinicians, basic scientists and engineers. This review first explores the prevalence, incidence, and morbidity of MBD and associated skeletal related events (SREs) for different cancer types. The relative benefits and risks of PT compared to after fracture care are discussed. The ability to predict SREs (specifically pathologic fractures) is then described for a number of different approaches, with validation using laboratory and clinical data. Clinical decision-making is then discussed in the context of patient survival, fracture risk assessment, and potential advantages and risks of PT.

PREVALENCE AND INCIDENCE OF METASTATIC BONE DISEASE

Prevalence of cancer cases in the United States is 1.2 million/year, but MBD prevalence is less clear. It has been estimated to be as high as 50% of all cancer cases (600,000 annually)¹ or as low as 25% (280,000 annually)². True prevalence is likely somewhere between those extremes. Prevalence is determined by input minus output, and the most important factor affecting prevalence is incidence. Incidence of MBD is strongly affected by type of primary carcinoma, predilection to metastasize, affinity of metastases for bone, and clinical ability to diagnosis bone metastases³. The most common carcinomas that spread to bone are prostate, breast, kidney, thyroid, and lung⁴ and 70% of MBD is accounted for by breast, prostate, and lung cancer². However, only 4.1%, 40.1%, and 10.3% of breast, lung, and prostate cancers present with stage IV disease, and not all of those have bone involvement. Bladder and melanoma also metastasize to bone in a relatively large percentage of cases (Table 1).

Table 1.

Percentage of primary carcinoma types with metastatic bone disease (MBD) and median survival from diagnosis of MBD.

Primary Carcinoma	Percentage with MBD	Median Survival After Dx of MBD (mo’s)
Breast cancer	65–75	19–25
Prostate cancer	65–75	12–23
Thyroid	60	48
Lung	30–40	6–7
Bladder	40	6–9
Renal	20–25	12
Melanoma	14–45	6

Adapted from Selvaggi et al.⁵.

The top five primaries can be remembered using the mnemonic Pb Ktl (or lead kettle), which also suggests the radiographic features of their metastases, prostate being osteoblastic, breast often mixed osteoblastic and lytic, and kidney, thyroid, and lung typically lytic. In terms of progression of disease, 12% of breast carcinoma patients who present with Stage I-III disease developed MBD within 60 months, and of the small fraction of Stage IV patients at diagnosis, 55% had MBD⁶. The most commonly diagnosed new cancers contrasted to the most commonly diagnosed cancers with stage IV disease at diagnosis are shown in Table 2⁷. Bone is among the top three sites of metastases, along with lung and liver^{8; 9}. Clinical evaluation likely underestimates MBD, as MBD at post-mortem examination usually exceeds clinical frequency^{9; 10}. This is consistent with the statement by Jaffe over fifty years ago, “70 per cent of patients who die of cancer are likely to have had bone metastases if the sampling of tissues at autopsy is adequate”¹¹.

Table 2.

Comparison of Most Common Overall Incidence of Primary Carcinomas Versus Most Common Metastatic Carcinomas in 2014 (United States).

Primary Carcinoma in Order of Most Commonly Diagnosed During 2014	Total Number of Cases Diagnosed During 2014	Primary Carcinoma in Order of Most Commonly Diagnosed with Metastatic Disease During 2014	Total Number of Stage IV Cases Diagnosed During 2014	Percentage of Stage IV Cases at Presentation (%)
Breast	232551	Lung/bronchus (non-small cell)	52153	40.1
Lung/bronchus (non-small cell)	129970	Pancreas	15670	45.2
Prostate	101879	Colon	14571	20.2
Colon	72241	Lung/bronchus (small cell)	13458	63.3
Melanoma of skin	54231	Non-Hodgkins Lymphoma (nodal)	12034	36.9
Urinary bladder	51005	Prostate	10528	10.3
Kidney or renal pelvis	46669	Breast	9459	4.1
Corpus uteri	43025	Kidney or renal pelvis	6504	13.9
Thyroid	38535	Urinary bladder	3892	7.6
Pancreas	34704	Rectum	3676	14.9
Non-Hodgkins Lymphoma (nodal)	32609	Thyroid	2915	7.6
Rectum	24628	Corpus uteri	2904	6.7
Lung/bronchus (small cell)	21266	Melanoma of skin	2415	4.5
Brain	20418	Brain	Unknown	Unknown

Derived from Site by Stage of Top 14 (out of 14) Sites Cancers Diagnosed in 2004–2014⁷.

Survival also affects prevalence. Improved survival means more patients with MBD remain alive at any given time⁴. Current treatments for breast, prostate, lung, and renal cancer have led to improved survival even after MBD diagnosis¹². Cumulative 1- and 3-year survival for breast cancer patients with MBD is 59% and 22%¹³. One- and 3-year survival for prostate cancer patients with MBD is 47 and 9%^{2; 14}. Primary cancer type affects prevalence because survival rates vary among primaries (Table 1). Finally, screening sensitivity affects prevalence^{14; 15}. In prostate cancer, the use of whole-body imaging and CT increased detection from 31.1% to 48.7% using PSA and bone scans¹⁶. Detection of MBD in lung carcinoma has increased with PET¹⁷.

Morbidity and Skeletal Related Events (SREs)

Morbidity of MBD may manifest as pain, pathologic fracture, decreased function in activities of daily living (ADLs), and overall decreased quality of life¹. Increased risk of fracture may present as impending pathologic fractures–weakened bone that has yet to fracture and pathologic fractures–after fracture through a weakened area of bone. Morbidity from MBD is often referred to by the term skeletal related events (SREs) and categorized into the specific subtypes: radiation therapy (RT), pathologic fracture (PF), bone surgery (BS), spinal cord compression (SCC), hypercalcemia (HC), and bone pain (BP). Factors affecting morbidity include anatomic location, type of lesion, and primary carcinoma (Table 3).

Table 3:

Skeletal related events (SREs) for breast, prostate, lung, renal cell, and thyroid cancer. Percent breakdown of SREs stratified by type.

				Distribution of Skeletal Related Event (SRE) Type (%)
Study	Metastatic Cancer	MBD Patient Sample (N)	Overall SRE (%)	Radiation Therapy	Pathological or Osteoporotic Fracture	Bone Surgery	Spinal Cord Compression	Hyper-calcemia
Jensen 201119	Breast	1494	48%	55%	19%	6%	20%	NA
Trinkhaus 201020	Breast	237	38%	79%	11%	NA	3%*	8%
Hussain 201821	Breast	3731	69%	32%	28%	8%	4%	NA
Kawai 201922	Castration resistant prostate cancer	2234	40.1%	27%	12%**	1%	2%	NA
McDougal 201623	Prostate	3297	43.5%	22%	8%	4%	7%	NA
Skov Dalgaard 201515	Lung	1668	67%	53%# 19%^	2%	1%	4%	NA
Woodward 201124	Renal cell	603	70%	61%	9%	12%	12%	6%
Farooki 201225	Thyroid	245	78%	46%^^ 44%+	16%^^ 19%+	19%^^ 25%+	16%^^ 10%+	3%^^ 2%+

^*Occurred prior to treatment with bisphosphonates;

^**includes traumatic and pathologic fractures;

^#Multiple radiation therapy treatments as the only SREs;

^{^}Radiation therapy as a solitary SRE;

^{^^}First SRE;

⁺Second SRE

The most common anatomic locations affected by MBD, include spine, pelvis, ribs, skull, proximal humerus, and the lower extremity long bones⁸. Types of SREs may be uniquely related to the involved anatomic site, such as SCC and spinal metastases⁹. However, BS, HC, PF, and RT, may occur at any anatomic site⁹. The most common anatomic locations requiring surgery in most series are the spine (68% in Nationwide Inpatient Sample, NIS), lower extremity (25% in NIS; especially the femur) and upper extremity (8.7% in NIS)¹⁸. In regards to need for femoral stabilization in this setting, Zickel and Mouradian in 1976¹¹ commented that, “The vast majority of lesions need no surgical treatment to maintain the bone’s function, but the fact that the femur is the bone that most often needs reinforcement or stabilization is noteworthy.”

Type of bone destruction has long been felt to have a direct relationship to morbidity, as osteolytic metastases that destroy bone (such as those from kidney, thyroid, lung cancer) are felt to pose higher risk for PF¹. In contrast, osteoblastic metastases (such as those from prostate and bladder cancer) pose less of a threat. Mixed lytic/blastic metastases (such as those of breast cancer) are thought to be of intermediate PF risk.

THE CASE FOR PROPHYLACTIC STABILIZATION

Before performing prophylactic stabilization, the surgeon must consider the relative benefits and risks of PT compared to after fracture (AF) treatment. In weighing evidence, one should consider how impending fractures are defined, since inclusion of patients not in need of PT may bias outcomes.

Overall and immediate post-operative survival has been suggested to be higher with PT compared to AF treatment (Table 4). Examination of the Nationwide Inpatient Sample (NIS) suggests operative treatment is protective of survival (p<0.001)¹⁸. Fracture of either the upper or lower extremity was a significant independent risk factor for mortality (p<0.001), the implication being that preventing fracture is important. However, there are subgroups of PT patients who may be at higher risk for shortened survival²⁶.

Table 4:

Published Literature Suggesting Survival Advantage to Prophylactic Stabilization.

	Patient Survival Prognosis
	Prophylactic (P)			Fracture (F)			P vs F
	Mean (mo’s)	Range (mo’s)	N	Mean (mo’s)	Range (mo’s)	N	P value
Zickel 197611^	13.8	1–48	11	4.7	0–5	35	--
Hardman 199227	15		27	2		34
Ward 200228	1 YR: 35% 2 YR: 19%	--	97	1 YR: 25% 2 YR: 10%	--	85	P=0.015
Arvinius 201429	14	1–34	21	11	1–49	44	N.S.
Philipp 201930	1 YR: 53% 2 YR: 40%		362	1 YR: 49% 2 YR: 31%		588	P = 0.018
	Immediate Post-op Death
	Prophylactic (P)			Fracture (F)			P vs F
	Percent	Deaths	N	Percent	Deaths	N	P value
Arvinius 201429	5%	1	21	11.4%	5	44	P=0.041

^{^}Using Zickel nail for prophylactic stabilization.

Potential clinical benefits of PT include pain and functional outcome improvements, lower estimated blood loss (EBL), and lower rates of transfusion and urinary tract infections (UTI). Controversy exists regarding improvements in pain and functional outcome. While some authors report earlier mobilization after PT (4.0 days PT vs 9.7 days AF, p=0.012), others found no difference^{29; 31}. “Improved ability to walk” has been reported in 78–100% of those undergoing PT compared to 58–76% of patients treated AF^{27; 29; 32}. In another series, support-free ambulation was achieved in 35% of PT patients compared to 12% treated AF²⁸.

Limited evidence suggests lower EBL and transfusion rate with PT. However, while differences are statistically significant, numbers are not large and therefore of debatable clinical importance (76,77). Absolute mean EBL difference reported in a single institution 12-year retrospective study was 200 cc (438 cc PT=97, 636 cc AF=85, p=0.01)²⁸. In a NIS study, rate of transfusion was 21.2% in PT patients compared to 28.6% in AF patients (p<0.001)³³ (Table 5). In the most comprehensive comparison to date of PT versus AF complications, a NIS study comparing 2750 PT patients to 2829 AF patients on multivariate analysis showed PT patients to have statistically higher rates of pulmonary embolism and deep venous thrombosis³³. A lower UTI rate advantage for PT did not hold up on multivariate analysis.

Table 5.

Comparison of Complications Between Prophylactic Treatment and Surgical Treatment after Pathologic Fracture in NIS

	Prophylactic fixation (%)	Pathologic fracture (%)	Univariate analysis		Multivariate analysis
			P value	Odds ratio	Adjusted odds ratio	P value
Pulmonary embolism	2.1	1.2	0.0008	1.79	2.1	0.0001
Deep venous thrombosis	3.2	2.2	0.03	1.43	1.5	0.03
Urinary tract infection	8.5	15.7	<0.0001	0.50	NS	NS
Blood transfusion	21.2	28.6	<0.0001	0.67	0.74	<0.0001

Adapted from Aneja et al.³³.

Another purported benefit of PT is lower cost. All series reporting discharge disposition favor PT compared to treatment AF (Table 6). With respect to LOS, the NIS showed no significant difference between PT and AF, while smaller retrospective single institutional series suggest benefits for PT. In the only comprehensive true cost analysis comparing PT to AF (2011–2014), PT showed lower direct and total cost, but this study consisted of only 40 patients (Table 7)³⁴. Implant costs were no different in that series.

Table 6.

Comparison of Rates of Discharge to Home and Hospital Stay Between Prophylactic Fixation and Pathologic Fracture Treatment.

	Number patients	Discharge rate to home			Hospital stay
		Prophylactic fixation (%)	Pathologic fracture (%)	P value	Prophylactic fixation (mean days)	Pathologic fracture (mean days)	P value
Aneja 201733	5579	58.4	30.2	<0.01	7.49	7.54	N.S.
Blank 201634	40	84	57	0.063	4 (all sites)	8	0.01
					4 (lower extremity)	7	0.053
Ward 200228	182	79	56	0.01	7	11	0.01
Arvinius 200229	65	NA	NA	NA	8	16	0.012
Hardman 199227	54	93	59	NA	NA	NA	NA

Table 7.

Cost Analysis of Prophylactic Fixation Compared to Pathologic Fractures.

	All anatomic sites			Lower extremities
	Prophylactic treatment (CU)	Pathologic fracture (CU)	P value	Prophylactic treatment (CU)	Pathologic fracture (CU)	P value
Implant cost	54 ±26	47 ± 25	N.S. (0.38)	52 ± 24	55 ± 27	N.S. (0.72)
Direct cost	227 ± 93	382 ± 300	0.038	237 ± 92	374 ± 272	0.06
Total cost	370 ± 171	642 ± 519	0.036	390 ± 169	634 ±477	0.053

Adapted from Blank et al³⁴.

APPROACHES TO ASSESS FRACTURE RISK

Classification/Scoring Systems for Long Bones with Metastatic Lesions

After considering the potential risks and benefits to prophylactic stabilization, the surgeon must consider approaches to assess fracture risk, as there are many. Early efforts to assess risk of impending fracture of long bones focused on clinical and radiographic risk factors using retrospective clinical series. Increasing pain, lesion size threshold (eg > 25 mm), osteolytic appearance, femoral location, lesion-to-bone width threshold (eg > 0.6), axial cortical involvement (eg > 30 mm), and circumferential cortical involvement (eg > 50%) either individually or in combination have been proposed as risk factors^{11; 35–39}. From these early papers, two widely quoted sets of criteria evolved -- Harrington’s criteria⁴⁰ and the “conventional definition” of impending pathologic fracture described by Carnesale⁴¹. Harrington’s criteria for prophylactic fixation of metastatic lesions of the femur included a lesion 25 mm or lager in size on radiographs, a lesion occupying more than 50% of the femoral cortex, and a lesion causing local pain with weight bearing. Carnesale similarly indicated that any lesion with significant pain and greater than 50% cortical involvement or larger than 30 mm represented increased risk of fracture. Interestingly, neither of these criteria has been assessed prospectively to evaluate accuracy.

Using the concepts of lesion location, lesion type (lytic/blastic), lesion size, and patient pain, Mirels’⁴² developed a widely referenced scoring system that combined radiographic and clinical risk factors into one score (Figure 1). The authors retrospectively applied this scoring scheme to a series of 78 lesions in 38 patients, and found the best combination of sensitivity and specificity to predict fracture was using a Mirels’ score of 9 or greater. With a Mirels’ score of 9, the probability of fracture was 33%, and this score was suggested as diagnostic for impending fracture. Overall, Mirels’ found high sensitivity and specificity (Table 8) with an overall accuracy of 90% in correctly predicting fracture. Pain and size were more predictive of clinical fracture compared to lesion (type) or lesion site. As an independent factor, lesion site did not appear to be a significant predictor of clinical fracture.

Figure 1.

Components of Mirels’ scoring for long bones and SINS for metastatic bone lesions.

Table 8:

Statistical measures of performance using Mirels’ and SINS scoring systems with long bone impending fracture predicted for a Mirels’ score of 9 or greater, and vertebral fracture.

Author	Scoring System	Region	Number of Patients	Fx | No Fx	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	Accuracy (%)
Mirels (1989)42	Mirels	Upper and lower extremity	38 (78 lesions)	27 | 51	81	94	88	91	90
Damron (2003)43	Mirels	Femur	12	2 | 5	91	35
Van der Linden (2004)45	Mirels	Femur	102 (110 lesions)	14 | 96	100	13	14	100	24
Nazarian (2015)47	Mirels	Femur	104	7 | 58	71	50	15	94	52
Goodheart (2015)46	Mirels	Femur	44	5 | 28	80	43	20	92	48
Bollen (2017)48	SINS	Vertebrae	110	94 | 16	45*	30	10	74

^*Average from four observers indicated.

Damron⁴³ reported that the Mirels’ rating system was reproducible, valid, and applicable across experience levels and training background. Reproducibility was high to very high for lesion type and site, but was fair for size and pain scores. As noted above, size and pain appear to be the most important predictors of fracture. In contrast, Howard et al⁴⁴, found only fair reproducibility for size and total Mirels’ score, and moderate reproducibility for site and lesion type. Subsequent application of Mirels’ scoring to independent patient series with metastatic lesions to long bones has been less promising (Table 8)^{43; 45–47}. While sensitivity of the method appears to be relatively strong (71–100%), the specificity is much poorer (13–50%) compared to the initial Mirels series. The high specificity found with the original Mirels manuscript may be due to the fact that the Mirels scheme was developed from one (training) data set, without independent application to a separate (test) data set. The large fraction of cases that were incorrectly predicted to fracture creates a clinical challenge, because if strictly followed, would result in a large population of unnecessary PT.

The independent predictive value of factors such as pain, lesion type and location in long bones has been questioned⁴⁵. Van der Linden⁴⁵ found that axial cortical involvement greater than 30mm (86/58% sensitivity/specificity) and circumferential involvement greater than 50% (43/82% sensitivity/specificity) were better predictors of fracture compared to Mirels (100/13% sensitivity/specificity). Reproducibility of the axial cortical metric was high, but quantifying circumferential involvement using multiple radiographic views was challenging. Strict use of the axial cortical metric would still lead to surgical overtreatment but to a lesser extent compared to Mirels’ or other conventional predictors such as pain and lesion type.

Three-dimensional analysis of lesional size and extent using CT imaging could improve size measures beyond that available from planar x-ray. Tatar et al⁴⁹ reported axial cortical involvement greater than 45mm (100/86% sensitivity/specificity) and circumferential cortical involvement greater than 30% (100/89% sensitivity/specificity) were strong predictors of long bone fracture. In addition to improvements in characterization of lesion geometry, the site of the lesion may have a large effect on bone strength. In a laboratory study, the femoral neck with an inferior-medial lesion was found to be less than half the strength of a femoral neck with a superior-lateral lesion of equivalent size⁵⁰. Therefore, there is likely potential for further improvement on geometry based fracture prediction by utilizing three dimensional imaging and data concerning lesional location beyond that initially conceived by Mirels.

Classification/Scoring of Spinal Instability

Given the high prevalence of metastatic lesions in vertebrae⁵¹, a classification system to diagnose impending spinal instability (Spinal Instability Neoplastic Score (SINS)) was developed in 2010⁵² by the Spine Oncology Study Group (SOSG). The goal was to develop a scoring system to guide clinicians in treating patients that may benefit from surgical consultation and to help guide surgical decisions. Some features overlapping with Mirels’ scoring system for long bones include lesion site, lesional nature (lytic/blastic/mixed), pain, and lesional extent (Figure 1). Additional spine components include spinal alignment and vertebral body collapse. The clinical classification is divided into stable, potentially unstable, and unstable groups.

Intra and inter-observer reliability has been reported to be very high for total SINS score⁵³. Interobserver reliability was fair for lesion type, moderate for alignment, vertebral body collapse, and posterolateral involvement, and almost perfect for location and pain. When compared to the binary gold standard (consensus opinion to identify stable and potentially unstable/unstable lesions), 96% sensitivity and 80% specificity was found for SINS assessment from 24 SOSG members. However, there is limited data on use of SINS as a predictive tool for progression of deformity or fracture. For a series of 110 patients with a median follow-up of 5.2 months, the ability to detect deterioration of spinal alignment requiring surgical stabilization or pathological fractures using SINS⁴⁸ was only modest (Table 8), with sensitivity and specificity of less than 50%.

Computed Tomography Based Structural Rigidity Analysis

Quantitative Computed tomography (QCT) scans provide full three-dimensional information about spatial geometry of bone and metastatic defects. In addition, bone density is available via the gray scale level of the bone, which correlates linearly with tissue mineral density. These two features (geometry and density) can be combined using engineering composite beam theory, resulting in measures of axial rigidity, bending rigidity, and torsional rigidity⁵⁴. Rigidity components can be calculated for each axial cross section of bone from a clinical QCT scan set. Because rigidity analysis does not calculate actual bone strength, the rigidity of homologous cross sections in the contralateral bone is also calculated, and the relative difference in rigidity is calculated. For cases that only involve one limb, density calibration with a phantom may not be needed using CTRA. If lesions are present in the contralateral limb, rigidity data from a normative library for the bone is used. An example of CT rigidity analysis (CTRA) for a 68 year-old male with a lytic lesion in the proximal left femur (Figure 2) illustrates the difference in rigidity based on axial position. A reduction of rigidity of 33% or more was considered to be at high risk of fracture⁴⁷.

Figure 2.

Computed tomography rigidity analysis (CTRA) of 68yo male with a Mirels’ Score of 10. CTRA indicates a significant risk of fracture at the site of the lesion and was prophylactically stabilized. In this case, axial rigidity was reduced by more than 33% at the cross section of the lesion, indicating a high risk of fracture.

CTRA has been performed on pediatric patients with benign skeletal lesions^{54; 55}, vertebral metastases⁵⁶, and metastatic lesions in the femur^{47; 57}. Overall, the approach appears to improve specificity without loss of sensitivity compared to the scoring methods described above (Table 9). One limitation with CTRA analysis is that the underlying beam theory assumes that the bones are straight. This may be appropriate for the diaphysis of long bone, but may introduce error at the ends of the bone that are often curved, such as the proximal femur, where many metastatic lesions occur. Curved beam theory (curved-CTRA) has been applied to the proximal femur and resulted in improved correlation between bending rigidity and failure load for curved-CTRA (r²=0.89) compared to straight beam CTRA (r² = 0.63)⁵⁸. This suggests that curved-CTRA could improve predictive capabilities of the general CTRA approach. One notable feature of the CTRA approach is that the calculation of rigidity on a slice-by-slice basis is not computationally expensive, and a complete analysis to estimate fracture risk can be conducted in 15 minutes⁵⁸.

Table 9:

Computed tomography rigidity analysis (CTRA) CTRA, Finite element analysis (FE), ¹⁸F-fluorodeoxyglucose (FDG) positron emission tomography/computed tomography (PET/CT), Parametric Response Mapping (PRM), and a weight scale analysis for clinical risk of fracture.

Author	Analysis	Lesion Type	Number of Patients	Fx | No Fx	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	Accuracy (%)
Snyder (2006)54	CTRA	Pediatric benign skeletal	36	18 | 18	100	78			89
Snyder (2009)56	CTRA	Vertebral metastases	94 (247 lesions)	11 | 236	100	55
Leong (2010)55	CTRA	Pediatric benign skeletal	102 (110 lesions)	0 | 41		97			97
Nazarian (2015)47	CTRA	Femur	104	7 | 58	100	90	54	100	91
Damron (2016)57	CTRA	Femur	94	6 | 88	100	60	15	100	63
Roth (2004)61	FE	Vertebrae	72 (92 vertebrae)	21 | 71	100	100
Goodheart (2015)46	FE	Femur^	44	5 | 28	80	86	50	96	87
Eggermont (2018)62	FE	Femur+	39	9 | 38	89	79	50	97	81
Sternheim (2018)63	FE	Femur+	11	0 | 11		63			63
Eggermont (2020)64	FE	Femur+	50	7 | 43	100	74	39	100	78
Ulaner (2017)65	FDG PET/CT	Femur	82	27 | 52	85	80	67	92	81
Hoff (2015)66	PRM	Vertebrae	5	5 | 5	100	71
Howard (2019)67	Weight Scale	Lower limb^^	59	26 | 33^^	97	65	79	86	81

^{^}level walking load condition.

⁺axial head load condition.

^{^^}estimated high fracture risk rather than actual fracture.

Finite Element Modeling

Application of the finite element (FE) method to predict fracture of the femur⁵⁹ and spine⁶⁰ with simulated metastatic lesions was first attempted in the 1990s. The FE approach is appealing because models recreate the three dimensional geometry of the bone and metastatic lesions (Figure 3) using Quantitative Computed Tomography (QCT) data, and each element has mechanical attributes (stiffness, strength, post yield behavior) based on local bone density. Loading conditions that reflect ADLs or loading from a fall can be applied to the model. While a definite load bearing capacity can be determined for the simulation, validation is needed to insure that the FE solution is accurate.

Figure 3.

CT based solid model of proximal femur (A) with a large intertrochanteric lytic defect (B). Elements are based on CT voxels (three dimensional pixels) (C) and material properties (elastic modulus shown here) of each element is based on CT mineral equivalent densities. Mechanical loading (D) can simulate different activities of daily living such as level walking or stair ascent. Reprinted from Goodheart et al⁴⁶.

Subject specific FE models can be developed from cadaver bones containing metastatic lesions or experimentally created lesions. The predicted strength from the FE analysis can be compared to the actual load bearing strength of cadaver bone for validation purposes (Table 10). Overall, correlation of FE predicted and actual bone strength is strong for the femur with clinical or experimentally created lesions (ave r²=0.87 for 6 studies, range 0.77–0.98)^68–73. The accuracy of FE predicted values is often expressed as the standard error of the estimate (SEE) normalized by the averaged measured strength (%SEE)⁷⁴. For the four non-linear finite element studies that report SEE for femurs with lesions, the average %SEE was 16.4%. These findings illustrate that when the loading is specifically defined, the FE approach can explain ~80% of the variability in load bearing strength, with errors in predicting actual failure magnitudes for a specific femur of 12–20%.

Table 10:

Correlation of finite element (FE) models with laboratory biomechanical testing. Standard error of the estimate (SEE).

Author	Tissue	Experimental Loading	Material Model	Number of Specimens	Lesions	Model-Exp Correlation (r2)	Mean Error (%)	SEE (N)	SEE/Mean (%)
Keyak (2005)84	Cadaver femur	Femoral head, stance phase	Non-linear, Density based strength, with post yield softening	26	5 with metastatic lesions	0.83	5.7	963	12.5
Keyak (2005)70	Cadaver femur	4 pt bend mid-diaphysis	Linear; Density based strength	12	8 with metastatic lesions	0.88		1380	13.6
Tanck (2009)72	Cadaver femur	Femoral head, stance phase	Non-linear, Density based strength, with post yield softening	10 (5 pairs)	Experimentally created peritrochanteric	0.92	18.7	1087	20.2
Derikx (2012)69	Cadaver femur	Femoral head, stance phase	Non-linear, Density based strength, with post yield softening	20 (10 pairs)	Experimentally created peritrochanteric, proximal diaphysis	0.92	15.7	1046	19.4
Yosibash (2014)73	Cadaver femur	Femoral head, stance phase	Linear; p-elements; Density based strength; surface strain limits	14	Metastatic lesions	0.78	23.3	3075	51.3
Benca (2019)68	Cadaver femur	Femoral head, 20 deg coronal plane	Nonlinear, density based strength, Hill criteria	32	Experimentally created, neck	0.77–0.98
Groenen (2018)76	Cadaver spine (2 motion segments)	Axial	Non linear, Density-based strength, softening, 4 material models	12	6 with experimentally created defects	0.22 – 0.25
Hojjat (2012)77	In vivo Rat vertebrae (2 motion segments)	Axial	Linear, density based strength; tissue strain correlation (10th percentile)	15	Inoculated breast cancer and prostate cancer cells	0.41

CT-based FE modeling has also been performed retrospectively for patients with femoral metastatic lesions prior to any fracture (Table 9). For example, Eggermont et al⁶⁴ performed subject specific FE analysis of 45 patients with femoral bone metastases using threshold of 7.5 x BW or lower identified as having high fracture risk. Using this criterion, the FE models were better at assessing fracture risk compared to clinical guidelines (lesions larger than 30mm) with improved sensitivity (100% vs 86%) and specificity (74% vs 42%). Similar sensitivity and specificity magnitudes were found with other FE models of clinical data sets (Table 9).

The correlation between FE modeling and experimental strength prediction for vertebral metastatic lesions have had mixed results. Using CT-based FE models of individual vertebral bodies with osteolytic lesions from multiple myeloma patients, cases with a vertebral fracture had 21% lower predicted yield force, and 27% lower work to yield compared to vertebrae that did not fracture⁷⁵. However, the ability to predict actual fracture strength has not been as promising. For FE models of human cadaver spines with experimentally created defects consisting of two motion segments (Figure 4), correlation between model and experiment subjected to axial loading was relatively weak (r²=0.22–0.25)⁷⁶. Predictive FE models for the spine has also been evaluated using a preclinical rat model with lytic and blastic vertebral lesions created using inoculated breast and prostate cancer cells⁷⁷. Even with use of microCT imaging, the correlation between axial load bearing from experiments with magnitude of high tissue strain in subject specific models was only moderate (r² = 0.41). Possible contributors to the poor strength prediction is the modified distribution of load between centrum and posterior elements during vertebral failure, the complex loading of the vertebral body via intervertebral discs, and endplate modeling that did not specifically model a very thin (<1mm) shell of bone. Multi-segment spine models with metastatic lesions are arguably much more complex from a modeling and experimental loading perspective compared to the proximal femur, and further development is necessary to improve the capability as a predictive tool.

Figure 4:

CT scan images of pre-experiment (A) and post axial load to failure (B) CT scans for three vertebral level construct. An experimental lesion was created for the middle vertebral body to simulate a metastatic lesion. White arrows indicate location of endplate fracture. Anterior (C) and lateral (D) views of specimen specific finite element models illustrate localization of failure to the middle vertebrae via plastic deformation (indicated in red/orange/yellow). Reprinted from Groenen et al.⁷⁶.

As an alternative to subject specific modeling of the spine, a biomechanically based regression model of spinal vertebral bulging⁶¹ with contributions from tumor volume, bone mineral density, pedicle involvement, disc degeneration, applied load, and loading rate was developed using a parametric FE approach⁷⁸. When applied to a series of 72 patients with 92 metastatic lesions, the authors found that sensitivity and specificity of 100% could be achieved for predicting fracture⁶¹. While promising, this work was retrospective and would need to be assessed in a prospective study to assess clinical utility. Nonetheless, the concept of combining a series of accessible factors that each contribute to vertebral fracture risk is appealing, and could be extended to other regions of the skeleton with metastatic lesions.

One challenge with respect to widespread use and implementation of QCT-FE to predict fracture risk is the effect of different CT imaging system and scanner settings on the calculated fracture load^79–82. For example, differences in predicted failure load of proximal cadaver femurs varied by 17% using CT scanners from three different manufacturers⁸². Note that the same femurs and calibration phantoms were used in all scanners. The choice of reconstruction kernel was found to have a large effect on bone mineral density (BMD) and FE predicted failure load. There are also differences in quantified bone mineral density, depending the amount and type of tissue surrounding the bone of interest. An in vivo study of femoral BMD for women using two different imaging systems, showed that there was a significant effect of body size on inter-scanner differences in BMD measures⁸¹.

There have been several recent advances in FE approaches that could improve the utility of QCT-FE based strength prediction. Density calibration without a phantom has been used for fracture risk assessment in patients with femoral bone metastases using QCT based FE models⁸³. No difference was seen in predicted failure load using phantom and phantom-less air-fat-muscle calibration. From a practical perspective, ability to develop bone FE models without need of a phantom allows access to a larger set of clinical scan data.

Material models used in metastatic lesion FE analysis have generally been derived from normal bone^84–87. However, there is limited data on density-modulus-strength relationships for metastatic bone and some indication that metastatic cortical bone has reduced modulus and compressive strength⁸⁸, even after accounting for altered bone density. Biphasic or poroelastic material models have been developed for osteolytic lesions where fluid is present within a porous solid matrix^{60; 89}. Blastic lesions, despite having greater apparent density, had lower modulus compared to normal bone⁹⁰. Specialized material models for blastic lesions that depart from normal dependence of strength and stiffness on bone density are therefore needed. Improvements in defining the geometry (edges) and intensity (density) in thin bone structures via image deblurring⁹¹ from clinical CT scan sets could also improve predictive capabilities of these models.

Other Fracture Prediction Approaches

¹⁸F-fluorodeoxyglucose (FDG) positron emission tomography/computed tomography (PET/CT) can be used to evaluate the metabolic activity of bone metastases. A retrospective study of metastatic breast cancer proximal femoral lesions showed that total lesion glycolosis (TLG) level could differentiate subsequent fracture patients (Table 9) with an accuracy of 81%⁶⁵. For this patient population, the results are comparable to CTRA and FE modeling.

Parametric Response Mapping (PRM) of CT Images uses a method in which changes in voxel gray scale (Hounsfield Unit) values are documented over time from longitudinal (sequential) CT scans⁶⁶. For a series of clinical cases with osteoblastic lesions, a rapid increase in local BMD was found to be predictive of impending fracture (100/71% sensitivity/specificity) (Table 9). However, sample size was very small and the developed PRM protocol should next be tested on a larger independent data set. One attractive feature of this approach is that it does not require complex FE modeling, but it does require image registration between CT scan sets.

Other simple tools, such as the ability to weight bear on the affected leg of patients with MBD through use of a weight scale has recently been proposed⁶⁷ as a method to assess fracture risk. Patients that could support less than 85% of body weight through the affected limb were found to be at increased risk of fracture. However, strict endpoints of pathologic fracture or no fracture were not used, but were based on clinical assessment of high or low fracture risk (Table 9).

CLINICAL DECISION-MAKING

In the shared decision-making discussion with the patient in the setting of possible impending pathologic fracture and in trying to come to the best decision with the patient, pros and cons must be weighed and discussed. As previously summarized, suggested potential advantages of PT, including survival, clinical benefits, and lower cost, should be considered. On balance, risks of PT, including early perioperative death and increased thromboembolic disease (TED) risk, must be weighed. Patient survival prediction following surgical treatment for MBD should be considered before surgery in order to determine if the patient’s predicted survival warrants operative intervention. Numerous survival prediction tools are available for MBD of both the spine and extremities^92–95. In two comparisons of 7–9 scoring systems for metastatic spine disease, the SORG nomogram showed highest accuracy for 30 (Area Under Curve, AUC 0.81) and 90 (AUC 0.78) day survival, but original Tokuhashi score was most accurate at 1 year (AUC 0.78)⁹⁴. The SORG nomogram is favored (94). Negative survival prognostic factors in the SORG nomogram include low hemoglobin, leukocytosis, older age, prior systemic treatment, lung/liver or brain mets, more than 1 mobile spine met, type 2 cancers, and European Cooperative Oncology Group (ECOG) 3 or 4 (describes patient functional status).

Patient survival tools for appendicular MBD include several models, scores, and nomograms. In direct comparison, the PATHFx model was most accurate at predicting 3 (AUC 0.70) and 6 (AUC 0.70) month survival, but the OPTIModel was most accurate for 12 (AUC 0.79) and 24 (AUC 0.77) month survival⁹⁶. PATHFx is available online (www.pathfx.org), whereas none of the other systems are to date. Factors associated with decreased survival include Type 2 or 3 tumors, visceral mets, femoral neck lesions, ASA > 3, ALP > 129 U/L, Albumin < 30 g/L.

In evaluation of three machine learning techniques to develop a survival prediction model, artificial neural network (ANN) had highest discrimination at 3 (AUC 0.89) and 12 (AUC 0.93) months compared to Bayesian belief network (BBN) and logistic regression⁹⁷. Decision (net benefit) analysis is a sophisticated decision analytic measure that provides information on balanced comparison of a test’s risks and benefits and has specific applicability to prediction tools⁹⁸. The authors compared ANN to BBN to regression applied decision analysis and demonstrated highest net benefit using ANN but suggested that BBN may be easier to implement in the clinical setting.

DISCUSSION

The global incidence rate of cancer is projected to increase to 20 million by 2030, and patients are also surviving longer with cancer and metastatic disease⁹⁹. For patients with MBD, identifying patients that have impending pathologic fractures is important because prophylactic treatment may reduce overall patient morbidity and mortality. However, correctly identifying patients truly at fracture risk remains a clinical challenge. Without good tools to identify impending fractures, potential remains for unnecessary surgery from a false positive test. Refined definitions of impending pathological fracture and spinal instability (through predictive tool assessment and validation) are therefore needed.

Of the numerous techniques available to predict fracture risk, the newer techniques that combine three-dimensional CT imaging and structural analysis (FE and CTRA) seem promising, but the clinical data sets used to evaluate these techniques are relatively small. In particular, there are few reported fracture cases, so that the confidence interval and uncertainty of the reported mean value for the sensitivity measure is very large. Access to a large number of pre-fracture CT scan sets, perhaps through use of phantomless density calibration⁸³, could be helpful in this regard.

Magnitude of load applied to a bone with metastatic lesions is likely a substantial contributor to fracture risk. For the proximal femur, loads from ADLs such as stance, gait, and stair climbing have been used in FE models, but more aggressive scenarios such as trip or fall have not been considered. Probability of encountering a range of loading activities could be assigned to individual patients as part of a probabilistic risk assessment tool¹⁰⁰ based on individual status (bedridden, ambulatory, at high risk of a fall). Combining this with a strength assessment (FE hip load capacity) could help guide decisions for PT.

In this era of “big data” and “machine learning”, we are on the cusp of many new developments. Machine learning, specifically decision tree (DT) and support vector machine (SVM), have been used to predict SREs, but access to truly “big data” is limited. Wang et al¹⁰¹ mined data for 1143 cancer patients, finding AUC 88.2% with only 4 variables for SRE prediction. Unfortunately, no analysis has been published yet for the subset of 248 pathologic fractures. It is this sort of truly big data with large numbers of pathologic fractures that may yield highly accurate fracture prediction. In a study combining CT and clinical data for 84 lung cancer patients with proximal femoral bone lesions, machine learning showed high accuracy (AUC 0.80) for fracture risk assessment, but 71% of patients fractured within 3 months, an unusually high level¹⁰².

Despite the promise of newer techniques (CTRA, FE, PET/CT, PRM, machine learning), they are not widely available and require specialized equipment/software and expertise to implement. Simpler screening/scoring tools described above (Mirels, cortical involvement, SINS) therefore still have an important role in clinical decision making. However, scoring prediction as applied to the femur suffers from poor specificity, and if followed strictly would greatly over-estimate the need for PT. SINS was only recently developed to predict spine instability, there is only one prospective study to date, and this has indicated sensitivity and specificity of less than 50%. Additional prospective assessment of these scoring tools is needed and application to larger data sets could also help refine important predictors of fracture.

Although the increasing number of recently published papers supporting PT over AF treatment is welcome, care must be taken in interpretation. “Impending fracture” definitions in those papers are often poor or absent. The potential is that misplaced benefit is assigned to PT procedures where patients would be better off without PT if not truly at risk for fracture. Future prospective data collection should provide detailed definitions and, when possible, use advanced fracture prediction tools.

The concept of combining CT imaging data and other clinical data could further improve fracture prediction. For example, Goodheart et al⁴⁶ showed improved prediction combining FEA and Mirels. CT-based descriptive variables combined with clinical features improved fracture prediction AUC from 0.62 to 0.80¹⁰². This approach of combining variables is similar to survival analysis and will require similar steps to validate.

Finally, current approaches do not consider the net benefit analysis for fracture risk testing. Net benefit analysis has been applied to compare survival prediction tools but not fracture prediction tools⁹⁷ in metastatic disease. Such application would be advantageous to determine the best fracture risk assessment tool from a broader perspective.

As a group, the specialists that evaluate and surgically treat patients with MBD and likely have the greatest potential to contribute to the advancement of care are orthopedic oncologists. As in other areas, evidence based medicine is needed to support the need for care of these patients by these specialists over community surgeons. Such efforts would surely be welcomed by those suffering from MBD.