Several prognostic indices have been devised to optimize patient selection for phase 1 oncology trials with no consensus as to the optimal score and none qualifying as a marker of treatment response.
Several prognostic indices have been devised to optimize patient selection for phase 1 oncology trials with no consensus as to the optimal score and none qualifying as a marker of treatment response.
Multivariate predictors of overall survival (OS) were tested on 118 referred patients to develop the Hammersmith Score (HS). The score's ability to predict OS, progression-free survival (PFS), and 90-day mortality (90DM) was compared with other prognostic indices. Changes in HS were recalculated during treatment.
Albumin < 35 g/L, lactate dehydrogenase > 450 U/L, and sodium < 135 mmol/L emerged as independent prognostic factors. These were used with equal weighting to devise the HS, a compound prognostic index ranging from 0 to 3. High (HS = 2-3) score predicted worse OS (hazard ratio [HR] = 6.5, P < .001), PFS (HR = 2.8, P = .01), and 90DM (OR = 9.0, P < .001). HS was a more accurate multivariate predictor of OS (HR = 6.4, P < .001, C-index = 0.72), PFS (HR = 2.7, P = .03), and 90DM (area under the ROC curve 0.703) compared with other scores. Worsening of the HS during treatment predicted for shorter OS (P < .001). HS retained prognostic and predictive ability following external validation.
HS is a simple, validated index to optimize patient selection and predict survival benefit from phase 1 oncology treatments. Prospective validation is ongoing. Cancer 2014;120:262–270. © 2013 American Cancer Society.
The main objectives of phase 1 trials in oncology are to define the toxicity profile and tolerability of novel compounds, explore their pharmacodynamic/pharmacokinetic parameters, and identify a maximum tolerated dose and a dose for use in future studies. Phase 1 trials are undertaken by patients with advanced cancers who have exhausted standard treatment options. Deciding which patients are suitable to enter a phase 1 study is challenging, because there is an inherent risk that they may experience toxicity from the experimental agent in exchange for limited clinical benefit. In order to maximize safety in this vulnerable patient population without jeopardizing trial success, recruitment relies on rigorous inclusion criteria such as adequate organ function, good performance status (PS), and predicted life expectancy of more than 3 months. The judiciousness of these criteria in guiding the selection of patients for phase 1 trials has been the subject of dispute in the past, because approximately one-third of the referred patients fail to fulfill these requirements at screening,[3, 4] and 15% to 20% die within 90 days of entering a study.[5-8] Selection for early-phase studies currently relies on the use of subjective parameters, such as clinician-assessed Eastern Cooperative Oncology Group (ECOG) or Karnofsky PS, which were originally developed for use in other clinical contexts. In an effort to improve patient selection, a number of prognostic indices specific to the phase 1 oncology patient population have been developed based on simple clinical parameters that emerged as independent predictors of survival from multivariate analyses (MVAs).[6, 8, 10-12]
The most widely evaluated algorithms include the Royal Marsden Hospital (RMH) score, which comprises serum albumin, number of metastatic sites, and lactate dehydrogenase (LDH)[10, 13]; the Prince Margaret Hospital Index (PMHI), consisting of albumin, number of metastatic sites, and PS; and the Nijmegen score, which comprises LDH, sodium, and hemoglobin (Hb). On the basis of these prognostic models, patients are stratified according to their mortality risk. Although these indices have been shown to predict 90-day mortality (90DM) and overall survival (OS) in the phase 1 oncology patient population, no consensus has yet been reached as to the optimal score to be used for patient selection. In addition, apart from the RMH score, the other scores lack independent validation, which limits their clinical applicability. Because the patients used to derive these scores were already participating in phase 1 oncology trials, the ability of these scores to prospectively predict those eligible for such studies is compromised. This is of particular significance if any claim is made that a given score assists clinicians in the eligibility assessment of potential participants to phase 1 trials.
Finally, none of these scores qualifies as a predictor of response to treatment. The recent shift away from cytotoxic chemotherapy and toward the evaluation of molecularly targeted agents (which are often cytostatic) in the phase 1 setting requires deployment of new specific and dynamic measures to predict or evaluate treatment benefit.[15, 16]
In th is study, we used a multivariate approach to identify variables that predicted survival in an unselected population of patients referred to our phase 1 oncology trials unit . The threshold values described in our study were based on conventionally employed laboratory thresholds as well as on previously published evidence.[8, 10, 11] These were used to devise a novel score named the Hammersmith Score (HS) , whose accuracy in estimating OS, PFS and 90DM was compared to that of existing prognostic indices and externally validated in an independently collected dataset from another trials center. We measured change to HS ( ΔHS) during phase 1 trial treatment and correlated this with survival to explore whether the HS could be used as a predictor of response to treatment.
Unselected referrals presenting to the phase 1 cancer trials program at the Wellcome Trust McMichael Clinical Research Facility ( MWT- CRF ), Hammersmith Hospital, Imperial College London, London, United Kingdom , between January 2007 and December 2011 were used to produce a training set . An independent data set including consecutive referrals to the University College London (UCL) Clinical Research Facility between April 2010 and January 2012 was used for validation . Complete demographic and clinical information were collected from medical r ecords . The RMH, PMHI, and Nijmegen scores were calculated for each individual as described before.[8, 10, 11] Clinical outcomes, including OS and 90DM, were calculated from the time of referral to our unit. For those who were still alive by the time of the data analysis, OS was calculated based on the time they attended for their last follow-up appointment. The cutoff date for the data analysis was May 16, 2012. For patients who were enrolled in a phase 1 trial, progression-free survival (PFS) was calculated from the date of the first dose to that of radiologically proven disease progression. Depending on protocol-specific requirements, computed tomography (CT) scan–based tumor assessment was carried out 6 to 8 weeks from study baseline. Response to treatment was assessed by a senior radiologist according to the Response Evaluation Criteria in Solid Tumors (RECIST), version 1.1. All patients gave their informed consent in accordance to the Declaration of Helsinki.
Pearson chi-square test was used to assess for associations between categorical variables. Univariate analysis of the different clinical factors associated with survival was carried out using Kaplan-Meier statistics and log-rank test. The independent prognostic value of each factor was explored using MVA according to Cox proportional hazard model using a stepwise backward selection approach. The SPSS statistical package version 19 (IBM SPSS Inc.) was used. Only variables displaying a significance level below 0.05 (2-sided) at univariate analysis were entered into the MVA, with factors displaying a corresponding P value ≥ 0.10 being removed from the Cox regression model. A logistic regression model was used to identify predictors for 90DM.
The receiver operating characteristic (ROC) curve method was used to measure the discrimination of 90DM by the different prognostic indices. The concordance index (C-index) method was used to rank scores according to their capacity of discriminating patients according to OS, with a value of 0.5 having no discriminative ability and a value of 1 having perfect discriminative ability. The effect of the candidate scores was assessed using the Cox model using R and SAS (SAS Institute, Cary, NC) software. The rms packages of Dr. Frank Harrell were used to identify a subset of predictors by backward elimination. For the assessment of prediction of multiple biomarkers, the C-index was adjusted within the rms package for the over-optimism produced by modeling and assessment being done on the same data via comparison with 150 bootstrap samples. All P values presented are 2-sided.
We identified 126 consecutive referrals for phase 1 trials to the MWT-CRF. Eight patients were excluded due to incomplete data and the final analyses were carried out using data from the remaining 118 patients, 85 of which (72%) had died at the time of analysis (Fig. 1). Baseline clinicopathologic characteristics of patients included in this study are summarized in Table 1. Median age was 63 years (range, 28-80 years) and the most prevalent tumor types were gynecological (35%) and gastrointestinal (33%).
|Characteristic||N (%)||Median (Range)|
|Age, y||63 (28-80)|
|Performance Status (ECOG)|
|No. of previous treatment lines||2 (0-8)|
|No. of involved areas||2 (0-5)|
|Locoregional disease only||10 (8)|
|1-2 distant metastatic sites||87 (74)|
|≥3 distant metastatic sites||21 (18)|
|Areas of metastatic spread|
|Extraregional lymphnodes||23 (20)|
|Other sites||21 (17)|
|Albumin, g/L||34 (13-43)|
|Serum LDH, IU/dL||249 (46-4218)|
|Sodium, mmol/L||138 (129-144)|
|Hemoglobin, g/L||11.5 (8.2-14.4)|
|White blood cell count, ×109/L||7.3 (2-12)|
|Platelet count, ×109/L||277 (69-626)|
|Primary tumor group|
|Gynecological cancers||42 (35)|
|Gastrointestinal cancers||39 (33)|
|Breast cancer||18 (15)|
|Genitourinary cancers||5 (4)|
|Lung cancer and mesothelioma||4 (3)|
|Skin cancers and melanoma||4 (3)|
|Head and neck||3 (2)|
The median OS for the entire cohort was 17 weeks (95% CI = 11-21) and the 90DM rate was 42% (95% CI = 33%-50%). Fifty-two patients (44%) were treated within 1 of the 7 phase 1 trials of which 5 were investigating molecularly targeted agents and 2 were investigating cytotoxic agents. No toxic death occurred. At the scheduled first tumor evaluation, 1 subject showed partial response (2%), 15 had stable disease (28%) whereas 36 had disease progression (70%). The clinical benefit rate (partial response+stable disease) was 37%. The median PFS in this subgroup was 7 weeks (95% CI = 5.6-8.6), whereas the median OS was 22 weeks (95% CI = 14-31). For those who did not enter a study, the median OS was 11 weeks (95% CI = 6-17).
Univariate analysis identified albumin < 35 g/dL, LDH > 450 IU/dL, sodium < 135 mmol/dL, ECOG PS ≥ 2, WBC > 10.5 × 109/L and a number of previous lines of chemotherapy ≥ 3, as significant negative prognostic factors for OS. In the MVA, albumin < 35 g/dL (P = .01), LDH > 450 IU/dL (P < .001), sodium < 135 mmol/dL (P = .06), and ECOG PS ≥ 2 (P = .04) were independent negative predictive factors for OS (Table 2). LDH > 450 IU/dL was identified as the only negative predictive factor for PFS in the univariate analysis (P = .004). All variables used in the MVA satisfied the proportional hazards assumption.
|Variable||N = 118||Overall Survival|
|Univariate Model||Multivariate Model|
|HR (95% CI)||Pa||HR (95% CI)||Pa|
|Age, y ≥65/<65||53/65||1.1 (0.7-1.7)||.68||–||–|
|Albumin, g/dL <35/≥35||44/70||2.4 (1.5-3.9)||<.001*||2.2 (1.7-4.1)||.01*|
|LDH, IU/dL >450/≤450||15/62||5.2 (2.5-10.7)||<.001*||4.5 (2-9.6)||<.001*|
|ECOG PS ≥2/0-1||29/86||4.5 (2.6-7.6)||<.001*||2.4 (1.0-5.7)||.04*|
|No. of metastatic sites ≥3/0-2||21/97||1.1 (0.6-1.9)||.78||–||–|
|No. of previous chemotherapy ≥3/0-2||53/63||1.7 (1.1-2.6)||.02*||–||–|
|Hemoglobin, g/dL <9/≥9||6/111||2.1 (1-5.4)||.1||–||–|
|Sodium, mmol/dL <135/≥135||21/92||3.2 (1.9-5.5)||<.001*||2.1 (1.0-6.0)||.06*|
|WBC, ×109/L >10.5/≤10.5||95/23||2.7 (2.5-4.7)||<.001*||–||–|
|Platelets, ×109/L >400/≤400||25/93||1.6 (0.1-2.8)||.06||–||–|
On the basis of the MVA, we developed a prognostic score consisting of 3 variables: albumin < 35 g/dL (1 point) versus albumin ≥ 35 g/dL (0 points); LDH > 450 IU/dL (1 point) versus LDH ≤ 450 IU/dL (0 points); sodium < 135 mmol/dL (1 point) versus sodium ≥ 135 mmol/dL (0 points). According to this index patients were stratified into a low-risk (0-1 points) and a high-risk (2-3 points) group, with those being in the low-risk group having a significantly longer OS (median OS of 31.2 weeks versus 8.9 weeks, P < .001) (Fig. 2, Table 3). The score was also able to predict PFS (median PFS of 8.7 weeks versus 4.3 weeks, P = .01) and 90DM (OR = 9.1, 95% CI = 2.8-28.9, P < .001). Patients with high-risk HS score at baseline were more likely to have poorer PS (P = .006), anemia (P = .01), presence of liver (P = .05) and lung metastases (P = .02) (Table 4). The normal values of the HS variables within our laboratory were: albumin ≥ 35 g/dL; sodium ≥ 135 mmol/dL and LDH ≤ 192 IU/dL.
|Median OS||95% CI||P||Median OS||95% CI||P||Median OS||95% CI||P||Median OS||95% CI||P|
|C-index||0.72||95% CI, 0.6-0.83||0.65||95% CI, 0.5-0.8||0.66||95% CI, 0.49-0.8||0.66||95% CI, 0.5-0.8|
|Variable||HS: 0-1||HS: 2-3||Pa|
|Age, <65 y/≥65 y||29/26||13/9||.61|
|ECOG PS, 0-1/≥2||48/5||13/8||.006*|
|No. of metastatic sites, <2/≥2||46/9||17/5||.52|
|Liver metastases, absent/present||28/27||6/16||.05*|
|Lung metastases, absent/present||40/15||10/12||.02*|
|Bone metastases, absent/present||43/12||19/3||.5|
|Hb, ≥12/<12 g/L||28/26||2/20||.001*|
For patients who were treated in a phase 1 study the HS was calculated at baseline and at the time of the first CT scan-based tumor assessment. According to the changes in their HS (ΔHS), patients were stratified in 2 groups: 1) those with a low-risk score (0-1) at baseline and after 2 cycles of treatment (n = 22); 2) those with a low-risk score (0-1) at baseline and a high-risk score (2-3) after 2 cycles of treatment (n = 9) or high-risk score at baseline and high-risk score after 2 cycles of treatment (n = 7). Those who maintained a low-risk score following treatment had a significant OS and PFS benefit compared to the other group (median OS 70.8 versus 17.8 weeks, P < .001; median PFS 9.4 versus 5.5 weeks, P = .01) (Fig. 2).
We derived the HS, RMHI, PMHI, and Nijmegen scores in our patient population and compared their performance for predicting OS, PFS, and 90DM. The results of the univariate and multivariate analyses are summarized in Table 5. All scores predicted OS in the univariate analysis but only HS (HR = 6.4, 95% CI = 3.2-13.2, P < .001) and PMHI (HR = 2.7, 95% CI = 1.4-5.2, P = .003) were identified as significant predictors of OS in MVA.
|Univariate Model||Multivariate Model||Univariate Model||Multivariate Model||Univariate Model||Multivariate Model|
|Score||N||HR (95% CI)||Pa||HR (95% CI)||Pa||HR (95% CI)||Pa||HR (95% CI)||Pa||OR (95% CI)||Pa||OR (95% CI)||Pa|
|HS 2-3/0-1||22/55||6.5 (3.4-13)||<.001*||6.4 (3.2-13.2)||<.001*||2.8 (1.2-6.4)||.01*||2.7 (1.0-6.8)||.03*||9.0 (2.8-28.9)||<.001*||9.0 (2.8-28.9)||<.001*|
|RMHI 2-3/0-1||42/35||2.1 (1.2-3.7)||.008*||–||–||2.1 (1.0-4.5)||.04*||–||–||1.7 (0.6-4.4)||.2||–||–|
|PMHI 2-3/0-1||32/42||2.9 (1.5-5.5)||.001*||2.7 (1.4-5.2)||.003*||1.7 (0.8-3.7)||.14||–||–||0.9 (0.4-2.5)||.9||–||–|
|Nijmegen 2-3/0-1||25/52||4.1 (2.3-7.6)||<.001*||–||–||2.4 (1.0-5.4)||.03*||–||–||5.2 (1.8-14.7)||.002*||–||–|
Prognostic models were ranked according to their discriminative ability in predicting OS by means of Harrell's concordance index. The C-index value was calculated for each prognostic system, as shown in Table 3. The HS ranked as the most accurate predictor of OS with a c-index value of 0.72 (95% CI = 0.60-0.83).
All prognostic scores were evaluated for predicting PFS, with the HS emerging as the only independent predictor (HR = 2.7, 95% CI = 1.0-6.8, P = .03). HS was also the most significant predictor for 90DM in the multivariate logistic regression (OR = 9.0, 95% CI = 2.8-28.9, P < .001) (Table 5) and ROC curve analysis (AUC = 0.703, P = 0.003) (Fig. 3).
The prognostic and predictive value of the HS was further assessed in an independent data set comprising 107 phase 1 patients with similar characteristics to those of the derivation set: median age 62 years (range, 20-81 years; P = .6); median OS 12.8 weeks (95% CI = 12.6-18.5, P = .2); ECOG PS (23% PS > 1; P = .16) and median number of previous treatment lines being 2 (range, 0-8; P = .34).
In the validation set the HS predicted for OS with those classified into the low-risk group (HS = 0-1) having a significant survival benefit against those with a high-risk score (HS = 2-3); median OS was 21 weeks compared to 7 weeks (P = .01). The HS was identified as a significant predictor of OS in the univariate analysis (HR = 23.4, 95% CI = 1.4-375, P = .02) and MVA (HR = 23.4, 95% CI = 1.4-375, P = .02).
For treated patients in the same data set, those who maintained a low-risk score after 2 cycles of treatment had a significant OS and PFS benefit compared to those in which HS worsened at the time of reassessment (median OS = 32.3 versus 13 weeks, P = .003; median PFS = 4.8 versus 1.9 weeks, P = .04).
Patient selection for phase 1 trials can be a challenging task as it involves the exposure of a vulnerable patient population with advanced cancer to potential toxicities from investigational agents in a setting where clinical benefit can be limited.[22, 23] There is a need for objective and evidence-based criteria to guide patient selection and on these grounds several prognostic scores have been developed.[6, 8-11] Among these, the RMH score has been prospectively validated as a significant predictor for OS and 90DM[13, 24, 25]; however, no consensus has yet been reached as to the optimal score to guide patient stratification and none is currently used in phase 1 protocol design.
In this study, we have shown that 3 simple, universally available laboratory parameters such as serum albumin, LDH, and sodium can be combined to generate an objective index that can be used to assist clinicians in selecting patients for early-phase oncology trials.
We have demonstrated the reproducibility of our newly qualified marker, the Hammersmith Score, in estimating survival by means of external validation in an independent patient cohort with similar characteristics, a necessary step before any given prognostic marker can be confidently used in the clinical setting.
In addition, we have explored the prognostic performance of the HS in comparison to that of other established scores including the RMH, PMH, and the Nijmegen scores. To our knowledge, this is the first study to comparatively test these indices in the broader population of phase 1 oncology referrals, whose clinical features are inherently different to those characterizing phase 1 trial participants.
In our analysis, when compared to other scores the HS was a better predictor of survival and early mortality, suggesting that it has greater prognostic power in the phase 1 patient population. Moreover, patients classified in the high-risk prognostic category had significantly poorer PS, lower hemoglobin levels and a higher proportion of visceral metastases, suggesting the ability of the HS to discriminate patients with a more aggressive disease phenotype.
Interestingly, in our analysis, some of the factors included in previously devised prognostic models did not retain prognostic value. This is particularly true for the number of metastatic sites that is included in the RMH score, which is one of the most widely accepted and externally validated prognostic algorithms. This discrepancy may have accounted for the relative underperformance of the RMH in our series. In addition, the original case series used to devise and validate the RMH score was only composed of patients already enrolled to phase 1 studies, who tend to be in better health than an unselected population of phase 1 referrals.
The HS was developed from the unbiased multivariate screening of prognostic traits, all of which have individually been shown to correlate with survival and disease progression in patients with advanced cancer. Reduced albumin serves as a marker of nutritional decline and has been closely associated with an ongoing inflammatory reaction that ultimately confers an adverse prognosis.[27, 28] Elevated serum LDH is known to reflect high tumor burden and impaired anticancer immune response, therefore strengthening the biological implications of this prognostic marker.
Hyponatremia is a common electrolyte disturbance with a multifactorial pathogenesis in cancer patients. Failure to maintain eunatremia has been shown to increase the risk of death, and this has been recently shown in the phase 1 cancer patient population. Although these factors have been previously tested as individual prognostic traits, this is the first time they are combined together to form a coherent prognostic index.
On the basis of previous knowledge showing that LDH and sodium may identify patients who are more likely to respond to targeted therapies, we also evaluated whether the HS could serve as a dynamic predictor of survival benefit from experimental targeted therapies, a function that has not been considered in the evaluation of other scores.
With our data, we provided evidence that patients remaining in a low-risk stratum (HS = 0-1) following 6 weeks of treatment have a survival gain of 53 weeks compared to those with a high-risk score (HS = 2-3) during treatment. These results, confirmed in the validation set, suggest that the clinical parameters used to devise the HS have the ability to reflect disease modulating effects from experimental therapies. Therefore, the HS could potentially be integrated as an additional endpoint alongside imaging to capture survival benefit in patients receiving investigational targeted therapies.
In our study, we also assessed the accuracy of other prognostic indices, some of which had never been externally validated. In particular, the PMH index ranked as the second most accurate predictor of survival, however, being based on PS, this could possibly lack objectiveness and reproducibility. Furthermore, based on our data, no independent prognostic value can be inferred for the Nijmegen score in our series.
Although the HS proved a better prognostic tool in our 2 unselected phase 1 populations than the RMH, PMH, and Nijmegen scores, the retrospective nature of our study as well as the relatively small size of both data sets should be acknowledged as limitations to our analysis. We therefore recognize the need to prospectively validate the HS in larger patient cohorts.
To conclude, our analysis shows that a simple prognostic index, the HS, is an externally validated score outperforming the other algorithms in predicting the survival of patients referred to a phase 1 clinic. Its reliability and objectiveness makes it a valuable clinical tool to improve the selection of trial participants. Furthermore, the dynamic changes of the HS during treatment make it a useful surrogate marker of response, predicting survival benefit from experimental therapies.
Based on these properties, the HS could therefore contribute to optimize the drug development process in its early phases, making patient stratification according to survival and response to treatment a clinically achievable aim.
We acknowledge support from the Imperial Experimental Cancer Medicine Centre (ECMC) and Biomedicine Research Centre grants from Cancer Research UK (CR-UK), NIHR, and UK Department of Health (DoH), as well as from the UCL ECMC and UCL BRC.
Drs. Seckl and Blagden acknowledge support from the Imperial ECMC and Biomedicine Research Centre grants from CR-UK, the NIHR and DoH. Martin Forster and Rebecca Kristeleit acknowledge support from the UCL ECMC and UCL BRC. All other authors made no disclosure.