Results of Literature Searches

The literature search identified 9,265 citations across all sources. Of these, 1,520 were obtained as full text. We identified 97 studies reported in 190 citations that met inclusion criteria.

Description of Included Studies

The included studies were published between 1991 and 2020. All studies reported on data collected using radiation therapy methods from 1990 or later to capture evidence that is relevant to today’s standard of care. Given that the included studies spanned a period of 30 years, we used meta-regressions to determine whether the reported effect sizes in newer studies tended to be larger than in older studies (because treatment effectiveness may have generally improved). We did not detect effects for all key outcomes that reported sufficient data (overall survival p=0.90, disease-free survival p=0.52, deaths due to brain metastases p=0.83, intracranial progression p=0.38, quality of life p=o.31, serious adverse events p=0.45, adverse events p=0.91, radiation necrosis p=0.71, headaches p= 0.95, fatigue p=0.91, seizure p=0.93, vomiting p=0.44). Hence, we did not pursue subgroup analyses for newer publications.

Half of the included studies had a unique trial identifier (the link to the study details can be found in the evidence table in Appendix D). A third of the included studies was based in the USA. The other studies were conducted in Australia, Austria, Canada, China, Egypt, France, Germany, Greece, India, Iran, Italy, Japan, Mexico, Netherlands, Poland, South Korea, Spain, Switzerland, Taiwan, and the UK. Six studies combined data from different countries. The large majority of studies (n=75) were RCTs, the remaining ones were cohort studies comparing two intervention cohorts. Study size varied from four participants included in an RCT that was closed early¹⁵⁸ to 3,536 participants included in a cohort study.⁹⁸

Most identified studies reported on WBRT as initial treatment and were relevant to Key Question 1. Twenty-four studies were relevant to Key Question 2. Only a dozen studies reported on post-surgery interventions (Key Question 3). With few exceptions, most studies contributed to Key Question 4 and reported on the presence or absence of at least one adverse event.

More than half of the included studies recruited patients with different primary tumor types, followed by studies in lung cancer patients, patients with melanoma, and patients with breast cancer. Similarly, the large majority of studies included patients with a range of prognoses. The number and volume of metastases was rarely described (see Appendix D).

Risk of Bias

The methodological quality of studies varied widely. Twenty-eight randomized studies reported adequate random sequence generation methods, with eighteen of them also describing allocation concealment. Seven studies state that a central office carried out the randomization, but did not provide the actual methods for randomization.^{102, 120, 131, 133, 141, 148, 163} Another thirty-five studies were described as randomized without further details. Twenty-four studies were classified as high risk of selection bias because they were not randomized. Three randomized studies were determined to have high risk of selection bias for our review.^{78, 89, 114} One study randomly assigned patients to treatment but compared outcomes to a historical group in their primary analyses, one randomized metastatic lesions but not the patients, and one did not report the randomization method and the treatment groups had important differences in baseline characteristics that could affect outcome.

Seven studies were described as double-blinded,^{82, 85, 100, 102, 120, 149, 168} although only two of those studies adequately reported their methods of blinding.^{102, 168} An additional study designated itself as single-blinded, but blinded both patients and examiners.¹⁶⁶ Patients were not blinded in 65 randomized studies, and therefore these studies had an unclear or moderate risk of performance bias. While not blinding participants probably did not affect the reliable determination of survival, it could have affected other outcomes such as neurocognitive endpoints and quality of life. Twenty-three studies were non-randomized observational studies, and therefore had higher risk of performance bias. One study was a non-randomized phase I study and therefor had higher risk of performance bias for the primary endpoints.¹³⁵

Thirty-four studies adequately reported attrition with no significant differences between treatment arms. Fifty-seven studies had moderate or unclear risk of attrition bias, mostly because attrition was unclear for some endpoints. Six studies had high risk of attrition bias, as they had significant attrition and/or attrition differed between treatment groups.^{81, 84, 94, 103, 105, 129}

Twenty studies had low risk of detection bias. In addition to the previously discussed double-blinded studies, twelve studies blinded the radiologists or neuropsychologists, or had a blinded review committee making assessments.^{76, 79–81, 84, 96, 107, 124, 127, 128, 155, 161} While the studies might be at risk for detection bias for some of their endpoints, we separated these studies from the remaining 53 without any blinding that were at moderate or unclear risk of detection bias. Twenty-three studies were non-randomized observational studies and were considered at higher risk of detection bias. A non-randomized phase I study was also determined to be at higher risk of detection bias.¹³⁵

Sixteen studies had low risk of reporting bias, with ten of these studies having their protocol readily available online or in a previous publication.^{10, 78–81, 106, 111, 115, 124, 164} In fifty-six studies the risk for reporting bias was unclear, mostly because the language that described which analyses were planned was not explicit, especially with subgroup and multivariable analyses. Twenty-five studies had high risk of reporting bias. Of those studies, 23 were observational studies. One randomized study did not report survival data and reported only significant results from their analyses.¹⁰⁷ Another randomized study did not report results from all of the outcomes collected and qualitatively reported some results.¹⁰³

Twenty-seven randomized studies analyzed their data by intent-to-treat and did not close early. Seventeen randomized studies had unclear risk of other biases, mostly because details were missing. Fifty-three studies had high risk of other biases. Of the interventional studies with high risk of bias, 25 were terminated early, either because of results during interim analyses or because of poor participant accrual (which was stated as a reason in 17 studies). One study was not analyzed by intent-to-treat¹⁵⁹ and another study’s modified intent-to-treat analysis was potentially problematic, as significant differences in exclusion were found between treatment arms.¹⁶⁴

Fifteen studies used quality of life assessments that were well-validated in brain metastases or brain tumor patients (e.g., FACT-Br, EORTC QLQ-C30 with BN20) or robust neurocognitive tests (e.g., HVLT-R). Forty-four studies did not assess quality of life or neurocognitive function. An additional seven studies collected data but did not completely report outcomes related to these endpoints.^{103, 104, 129, 135, 154, 165} The remaining twenty-three studies with moderate rating did not specify the assessment tool used, used an assessment validated in other disease settings, or used cognitive tests such as the MMSE that are used in the assessment of dementia. Eight studies used performance scales (e.g. ECOG performance scale) with only one measure usually assessed subjectively by a clinician,^{76, 97, 101, 122, 136, 147, 156, 160} and are therefore considered more problematic in assessing quality of life or function.

Eleven studies were determined to have low risk of overall bias for effectiveness outcomes.^{10, 79, 80, 85, 90, 99, 100, 124, 149, 155, 168} Fifty-three had moderate or unclear risk of bias, with 18 of those studies not having enough details for assessment. The remaining were considered high risk of bias.

Nine studies collected adverse event data systematically and prospectively.^{10, 79–81, 90, 97, 111, 116, 135} Sixty prospective studies reported adverse events, but it was unclear how events were collected. Of the studies rated as high risk in their collection of adverse events, 17 did not report adverse events or simply stated that no adverse events occurred. The remaining studies either collected their events retrospectively, or collected only specific events (e.g., radionecrosis or surgical complications). Twenty-nine studies reported adverse events rigorously, including severity and a variety of adverse events by treatment groups. Twenty-three studies reported adverse events for only a limited number of non-hematological events. Of the 45 studies rated as high risk, 17 did not report adverse events. The remaining studies reported adverse events for their whole cohort but not by treatment arm, did not report rates of events, or reported only on specific adverse events (e.g., radionecrosis). Taking into consideration the method of collection and reporting of adverse events, 22 studies were considered relatively low risk, 29 were considered moderate or unclear, and 46 were considered high risk in their adverse event assessment.

Details on Strength of Evidence

We used the criteria outlined in Appendix A to assess the strength of the body of evidence for each Key Question. All findings started at high strength of evidence as the results were mostly based on RCTs. We did not upgrade any findings. Most often we downgraded results due to imprecision, study limitations, or indirect evidence. The reasons for downgrading are included in the summary of findings tables.