A Model for Predicting Day-100 Stem Cell Transplant-Related Mortality in AL Amyloidosis

Abstract

Autologous stem cell Transplant (ASCT)-related mortality (TRM) in AL amyloidosis remains elevated. AL amyloidosis patients (n=1718) from 9 centers, transplanted 2003–2020 were included. Pre-ASCT variables of interest were assessed for association with Day-100 all-cause mortality. A random forest (RF) classifier with 10-fold cross-validation assisted in variable selection. The final model was fitted using logistic regression. The median age at ASCT was 58 years. Day-100 TRM occurred in 75 patients (4.4%) with the predominant causes being shock, high-grade arrhythmia, and organ failure. Ten factors were associated with Day-100 TRM on univariate analysis. RF classifier using these variables generated a model with an area under the curve (AUC) of 0.72±0.12. To refine the model selection using importance hierarchy function, a 4-variable model [NT-proBNP/BNP, serum albumin, ECOG performance status (PS), and systolic blood pressure] was built with an AUC of 0.70±0.12. Based on logistic regression coefficients, ECOG PS 2/3 was assigned two points while other adverse predictors 1-point each. The model score range was 0–5, with a Day-100 TRM of 0.46%, 3.2%, 5.8%, and 14.5% for 0, 1, 2, and ≥3 points, respectively.

This model to predict Day-100 TRM in AL amyloidosis allows better-informed decision-making in this heterogeneous disease.

Introduction

Autologous stem cell transplantation (ASCT) is an effective therapy for AL amyloidosis leading to a high overall complete and partial response rate, and durable remissions.¹ However, high-dose chemotherapy is associated with high morbidity and can be offered only to a subset of AL amyloidosis patients who are considered fit for this therapy. Improvement in patient selection has decreased Day-100 transplant-related mortality (Day-100 TRM) from 13–24% in the early years to a rate of 3–5% in high-volume transplant centers.^2–10 TRM following ASCT for AL amyloidosis is primarily affected by cardiac status,^{8, 11} although other factors influence it such as renal failure,¹² autonomic nerve involvement,¹³ and performance status (PS).⁷

With the emergence of more effective standard-intensity therapies, particularly bortezomib, daratumumab, and venetoclax, the feasibility of attaining deep hematological responses, including hematological complete response, without significant morbidity and mortality is re-shaping the role of ASCT in this disease.^14–16 The question of patient safety remains relevant considering the improvements in disease management and alternate therapeutic options. However, tools to accurately predict Day-100 TRM are limited. This study built a model that provides an estimated risk of Day-100 TRM in transplant-eligible AL amyloidosis patients using clinical factors identified through statistical analysis.

Methods

The institutional review board approved the study at all centers. All patients with AL amyloidosis who underwent ASCT for AL amyloidosis between January 1, 2003, and December 31, 2020, were included. To be eligible, patients must have had amyloid typing (using mass spectrometry, immunohistochemistry, or immunofluorescence) confirming immunoglobulin light chain type. Patients had to have at least 3 months follow-up from Day zero of ASCT (the day of stem cell infusion) unless earlier death occurred. All-cause mortality within 100 days of Day zero was considered TRM, irrespective of the cause of death.

Data were collected from the time of diagnosis, prior therapies including induction therapy (if given), pre-transplant testing, transplant-related data, and survival. The pre-transplant evaluation data was obtained before stem cell mobilization. Blood and urine studies were obtained within 1-month prior to transplant, while echocardiogram and pulmonary function testing were included if performed within 3 months prior to transplant. The estimated glomerular filtration rate (eGFR) was calculated using the MDRD formulation.

Except for ECOG performance status, the proposed model was developed using categorical variables in a binary dichotomization for ease of use. The cutoffs used for dichotomization were derived from existing literature on AL amyloidosis (or stem cell transplantation when data in AL amyloidosis literature was not available), to allow the identification of patients at higher risk of death. Considering the exclusion criteria for ASCT in AL amyloidosis,¹ these cutoffs were chosen differently than the exclusion criteria for transplantation cutoffs. These cutoffs include dFLC ≥180 mg/L,¹⁷ systolic blood pressure <100 mmHg,¹⁸ interventricular septum >16 mm,¹⁹ left ventricular ejection fraction <50%,²⁰ cDLCO <65% of predicted,²⁰ FEV1 <65% of predicted,²⁰ NT-proBNP/BNP ≥1800/400 pg/mL,¹⁷ serum albumin <2.5 g/dL,²¹ eGFR <30 mL/min,²² and alkaline phosphatase ≥2 folds the upper limit of normal.²³ Troponin could not be included in our modeling given the multiple assays used across institutions (troponin T versus I and classical versus high sensitivity assays) which could not be reconciled. The European modification of the Mayo 2004 cardiac stage at ASCT,^{24, 25} when used, was restricted to patients with troponin T or high-sensitivity Troponin T measurements, with Troponin T and high-sensitivity Troponin T harmonized as previously reported.²⁶ ECOG performance status (PS) was used as a categorical 3-level variable, given distinct Day-100 TRM in each ECOG PS category (1.5%, 5%, and 13.1% for ECOG 0, 1, and 2/3, respectively).

Statistical analysis

Baseline characteristics of the patient cohort were summarized using descriptive statistics, with continuous variables reported as medians with interquartile ranges (IQRs) and categorical variables summarized as frequencies and percentages. Missing data were imputed using the Multiple Imputation by Chained Equations (MICE) method.²⁷ Two notable exceptions for data imputation were dFLC and NT-proBNP/BNP imputations, where we first employed informed imputations as detailed below, followed by the MICE method. For patients with missing pre-ASCT dFLC who did not receive induction, baseline dFLC was used to replace missing pre-ASCT dFLC when available (n=26). In patients with missing pre-ASCT NT-proBNP/BNP and without heart involvement per investigator assessment (i.e., normal echocardiography and lack of cardiac symptomatology), we assigned non-adverse pre-ASCT NT-proBNP/BNP for such patients (n=104). Overall, data imputation was performed for the following variables (in decreasing order): FEV1 (n=155, 9%), NT-proBNP/BNP (n=153, 8.9%), cDLCO (n=125, 7.3%), IVS (n=85, 4.9%), EF (n=46, 2.7%), dFLC (n=38, 2.2%), ECOG score (n=25, 1.5%) and SBP (n=1, <0.1%).

Univariate logistic regression analyses were conducted to identify potential predictors of Day-100 TRM, with associations summarized using odds ratios (ORs) and 95% confidence intervals (CIs). A random forest (RF) classifier with tenfold cross-validation was used to rank the importance of pre-ASCT variables, refining the selection of predictors for the final multivariable logistic regression model. The final model variables were selected based on statistical performance and clinical relevance. The model’s performance was evaluated using the area under the receiver operating characteristic curve (AUC-ROC) with 95% CIs. A scoring system was developed based on logistic regression coefficients, with each predictor assigned a score proportional to its regression coefficient. The model resulted in a score range of 0 to 5, stratifying the risk of Day-100 TRM into four categories (0, 1, 2, and ≥3 points). Comparison of the TRM model with the cardiac stage model for Day-100 TRM prediction was performed using DeLong’s test for two correlated ROC curves. Statistical analyses were conducted using SAS version 9.4M5.1 and Python version 3.11.5.

Results

Baseline characteristics

1718 patients from 9 centers are included. Their baseline characteristics are listed in Table 1. Most patients had 1 (43.0%) or two (33.3%) involved organs, with the dominant organs being the kidneys (68.7%) and heart (51.5%). Early Mayo 2004 cardiac stage was assigned to most patients at diagnosis (stage I and stage II 39.3% and 43.7%, respectively). Most patients (93.6%) underwent ASCT for newly diagnosed disease. Among the newly diagnosed patients, the median time from diagnosis to ASCT was 4.9 months (IQR 3.3–7.8). Induction therapy was given to 51.2% of patients for a median duration of 3 months (IQR 2–5).

Table 1.

Patients’ baseline characteristics

	Whole cohort (n=1718)
Age in years, median (IQR)
At diagnosis	58 (52–63)
At ASCT	58 (53–64)
Female sex, n (%)	678 (39.5%)
Lambda Light chain isotype, n (%)	1292 (75.2%)
Heavy chain isotype, n (%)
IgG	610 (35.5%)
IgA	194 (11.3%)
IgM	66 (3.8%)
IgD	19 (1.1%)
No heavy chain association	829 (48.3%)
dFLC at diagnosis, median, mg/L (IQR)	142 (54–454)
BMPCs at diagnosis, median, % (IQR)	10 (5–15)
No. of involved organs, %
1	739 (43.1%)
2	571 (33.2%)
>2	408 (23.7%)
Organ involvement pattern, n (%)
Kidneys	1193 (69.4%)
Heart	885 (51.5%)
GI	335 (19.5%)
Liver	211 (12.3%)
Peripheral nerve	218 (12.7%)
Autonomic nerve	194 (11.3%)
Other	232 (13.5%)
Mayo 2004 cardiac stage, I/II/IIIA/IIIB, %	39.3/43.7/13.6/3.4

Abbreviations: ASCT, Autologous stem cell transplantation; BMPCs, Bone marrow plasma cells; dFLC, difference between involved to uninvolved light chains;

IQR, interquartile range, GI, Gastrointestinal

Pre-transplant evaluation and transplant data

Pre-ASCT ECOG PS was 0, 1, and 2/3 in 35.9%, 57.0%, and 7.1% of patients, respectively. ECOG PS 2/3 was significantly more common in the following subgroups: cardiac involvement (10.2% vs. 3.8% in those without heart involvement, P<0.001), autonomic nerve involvement (13.9% vs. 6.2%, P<0.001), renal stage III (14.5% vs. 6.2% in renal stage I/II, P<0.001), cardiac stage IIIB, IIIA, and II (17.5%, 10.8%, and 8.9%, respectively, vs 3.5% in stage I; P<0.001) and SBP <100 mmHg (13.2% vs 6.3%, P<0.001).

Pre-ASCT dFLC <180 mg/L was present in most patients (74.2%), and these patients were more likely to have received induction than patients with dFLC ≥180 mg/L (59.1% vs 38.4%,; P<0.001). Other pre-transplant variables are summarized in Table 2. Adverse features were present in 6.9–25.8% of patients.

Table 2.

Pre-ASCT evaluation

	Whole cohort (n=1718)
ECOG performance score 0/1/2–3, %	35/58/7
SBP, mmHg
Median, IQR	116 (106–128)
<100 mmHg, n (%)	213 (12.4%)
dFLC, mg/L
Median (IQR)	62 (16–186)
≥180 mg/L	443 (25.8%)
Left ventricular ejection fraction (LVEF), %
Median (IQR)	62 (56–66)
LVEF <50%, n (%)	118 (6.9%)
IVS, mm
Median (IQR)	12 (10–14)
IVS >16 mm, n (%)	211 (12.3%)
cDLCO, % of predicted
Median (IQR)	80 (69–91)
cDLCO <65% of predicted, n (%)	306 (17.8%)
FEV1, % of predicted
Median (IQR)	88 (77–102)
FEV1 <65% of predicted, n (%)	155 (9.0%)
NT-proBNP/BNP, pg/mL
NT-proBNP, median (IQR)	499 (163–1679)
BNP, median (IQR)	130 (51–292)
≥1800/400 pg/mL, n (%)	350 (20.4%)
Serum albumin, g/dL
Median (IQR)	3.4 (2.6–4.0)
<2.5 g/dL, n (%)	322 (18.7%)
eGFR, mL/min/1.73 m2
Median (IQR)	73 (53–92)
<30 mL/min/1.73 m2, n (%)	171 (10.0%)
Alkaline phosphatase, folds of ULN
Median (IQR)	0.7 (0.5–1.0)
≥x2 folds of ULN	149 (8.7%)

Abbreviations: ASCT, Autologous stem cell transplantation; BNP, B-type natriuretic peptide; cDLCO, corrected diffusion capacity for carbon monoxide; dFLC, difference between involved to uninvolved light chains; ECOG, Eastern Cooperative Oncology Group; eGFR, estimated glomerular filtration rate; FEV1, forced expiratory volume in 1 second; IQR, interquartile range; IVS, Interventricular septum; LVEF, Left ventricular ejection fraction; NT-proBNP, N-terminal of pro b-type natriuretic peptide; SBP, Systolic blood pressure; ULN, Upper limit of Normal

Nearly all patients received melphalan conditioning (n=1704, 99.2%) with the remaining patients conditioned with carmustine, etoposide, cytarabine, and melphalan (BEAM) conditioning. Of the melphalan-conditioned patients, 63.3% received a melphalan dose of 200 mg/m2.

The Day-100 TRM rate was 4.4% (n=75). It was significantly higher in 2003–2011 compared to 2012–2020 (5.5% vs 3.2%, P=0.02). The three main causes for Day-100 TRM were shock (septic, hypovolemic, or cardiogenic; n=27), high-grade arrhythmia (n=26), and organ failure (n=22). The timing of death in these 75 patients is provided in Figure 1. Weeks 1–3 had the highest rates of TRM, with a peak in week 2 (n=17, 22.7% of all Day-100 TRM). Subsequent peaks were noted in weeks 6 and 7 (n=8 and n=6, respectively) and week 13 (n=7). Causes of death differ between the 2003–2011 and 2012–2020 periods. In the 2003–2011 period, the main cause of death was organ failure (n=20), while shock and high-grade arrhythmias were documented as the cause of death in 14 patients each. In the 2012–2020 period, organ failure significantly declined as a cause of death (n=2) while high-grade arrhythmias (n=14) and shock (n=13) were the leading causes of day-100 death (P=0.0035)

Figure 1:

Distribution of transplant-related deaths by week from day 0 till Day-100 (week 15).

Univariate analysis for predictors of Day-100 TRM

Sex and light chain isotype did not predict Day-100 TRM in univariate analysis (Table 3). An additional 12 pre-ASCT variables were assessed for association with Day-100 TRM; this analysis results are presented in Table 3. Except for age at ASCT and eGFR, all other variables were significantly associated with Day-100 TRM on univariate analysis. The odds ratio of these variables ranged from 1.6 to 10.2. Serum albumin <2.5 g/dL was almost exclusively seen in patients with renal involvement (99% of patients with serum albumin <2.5 g/dL had renal involvement) and was highly correlated with 24-hour proteinuria (median 24-h proteinuria 9.3 vs. 1.5 g for patients with serum albumin <2.5 vs ≥2.5 g/dL, respectively; P<0.001).

Table 3.

Univariate analysis for Day-100 TRM prediction

Predictor	N (%)	OR (95% CI)	P-Value
Male sex	1040 (60.5%)	1.4 (0.9–2.3)	0.178
Lambda light chain isotype	1292 (75.2%)	0.7 (0.4–1.2)	0.23
Age at ASCT ≥ 60 years	780 (45.4%)	1.5 (0.9–2.3)	0.1
Pre-ASCT dFLC ≥180 mg/L	443 (25.8%)	2.50 (1.57–4.00)	<0.001
ECOG 1	979 (57.0%)	3.64 (1.78–7.45)	<0.001
ECOG 2/3	122 (7.1%)	10.20 (4.39–23.68)	<0.001
LVEF <50%	118 (6.9 %)	3.40 (1.84–6.27)	<0.001
IVS >16 mm	211 (12.3%)	2.37 (1.37–4.12)	0.002
cDLCO ≤65% of predicted	306 (17.8%)	2.27 (1.38–3.75)	0.001
FEV1 ≤65% of predicted	155 (9.0%)	2.93 (1.64–5.24)	<0.001
Pre-ASCT NT-proBNP/BNP ≥1800/ ≥400 pg/mL	350 (20.4%)	3.29 (2.05–5.27)	<0.001
eGFR<30 mL/min/m2	171 (10.0%)	1.59 (0.82–3.08)	0.167
Serum albumin <2.5 g/dL	322 (18.7%)	1.99 (1.20–3.30)	0.008
Alkaline phosphatase >X2 ULN	149 (8.7%)	2.09 (1.10–3.98)	0.024
SBP <100 mmHg	213 (12.4%)	2.73 (1.60–4.66)	<0.001

Please note that all predictors are assessed at the pre-ASCT stage

Abbreviations: ASCT, Autologous stem cell transplantation; BNP, B-type natriuretic peptide; cDLCO, corrected diffusion capacity for carbon monoxide; CI, Confidence interval; dFLC, difference between involved to uninvolved light chains; ECOG, Eastern Cooperative Oncology Group; eGFR, estimated glomerular filtration rate; FEV1, forced expiratory volume in 1 second; IVS, Interventricular septum; LVEF, Left ventricular ejection fraction; NT-proBNP, N-terminal of pro b-type natriuretic peptide; OR, Odds ration; SBP, Systolic blood pressure; TRM, Transplant-related mortality; ULN, Upper limit of Normal

Bold indicates P<0.05

Building a predictive regression model for Day-100 TRM

The 10 significantly associated pre-ASCT predictors of Day-100 TRM on univariate analysis were entered into a Random Forest model, an ensemble bagging method. To ensure robust performance evaluation and minimize overfitting, 10-fold cross-validation was employed. The logistic regression model output is presented in the Supplementary Table. The ROC-AUC of this model was 0.72 ± 0.12. Using SHAP (SHapley Additive exPlanations) analysis to assess the impact of each variable on model prediction, variable importance hierarchy is depicted in Figure 2. To create a simpler model for clinical application at the point of care, we focused further on the top 6 variables in the SHAP analysis, which included in decreasing order of importance: ECOG PS, NT-proBNP/BNP, dFLC, serum albumin, SBP and cDLCO. A model featuring these 6 variables was then evaluated based on statistical performance, clinical relevance, and ease of application. dFLC, once an important prognostic factor in AL amyloidosis, is less meaningful since the introduction of anti-CD38 monoclonalantibodies y^{28, 29} and was removed. cDLCO did not show statistical significance for Day-100 TRM prediction in multivariable models and was also removed. Therefore, our final model includes 4 predictors: ECOG PS (adverse being 1 or 2/3), serum albumin (<2.5 g/dL), NT-proBNP/BNP (≥1800/400 pg/mL), and SBP (<100 mmHg). This model has an ROC-AUC of 0.70 ± 0.12 (Figure 3). ORs for risk of Day-100 TRM are summarized in Table 4. Based on similar coefficient weight and confidence intervals for these variables, each factor was assigned zero when an adverse feature was absent and a score of 1 when was present. ECOG PS 2/3 was assigned a score of 2, given a higher coefficient for Day-100 TRM. The model’s total score ranges from 0 to 5.

Figure 2:

SHAP summary plot. The plot depicts the importance of various clinical features in predicting the risk of TRM, as determined by the sum of absolute SHAP (SHapley Additive exPlanations) values. The features are ranked in descending order of importance, with higher SHAP values indicating a greater influence on the model’s ability to predict TRM.

Abbreviations: ASCT, Autologous stem cell transplantation; BNP, B-type natriuretic peptide; cDLCO, corrected diffusion capacity for carbon monoxide; dFLC, difference between involved to uninvolved light chains; ECOG PS, Eastern cooperative group performance status; FEV1, forced expiratory volume in 1 second; IVS, Interventricular septum; LVEF, Left ventricular ejection fraction; NT-proBNP, N-terminal of pro b-type natriuretic peptide; SBP, Systolic blood pressure; ULN, Upper limit of Normal

Figure 3:

Receiver operating characteristic curve for the prediction of Day-100 TRM using a 4-predictor model.

Table 4.

Four-factor multivariable regression model for Day-100 TRM prediction

Predictor	OR (95% CI)	P-Value
Pre-ASCT NT-proBNP/BNP ≥1800/ ≥400 pg/mL	2.53 (1.54–4.15)	<0.001
Serum albumin <2.5 g/dL	1.89 (1.11–3.21)	0.018
ECOG 1	2.97 (1.44–6.14)	0.003
ECOG 2/3	6.45 (2.70–15.43)	<0.001
SBP <100 mmHg	1.90 (1.09–3.31)	0.024

Abbreviations: BNP, B-type natriuretic peptide; CI, Confidence interval; ECOG, Eastern Cooperative Oncology Group; NT-proBNP, N-terminal of pro b-type natriuretic peptide; OR, Odds ration; SBP, Systolic blood pressure; TRM, Transplant-related mortality

Bold indicates P<0.05

We then applied the model scoring to our population. Score distribution was as follows (Table 5): 0 (n=431, 25.1% of patients), 1 (n=688,40.1%), 2 (n=413, 24.0%), and three and above (n=186, 10.8%). The respective Day-100 TRM rates were 0.46%, 3.2%, 5.8%, and 14.5%, respectively. The rates for the 2003–2011 period were 1.1%, 3.3%, 6.4%, and 19.2%, while for the period 2012–2020, 0%, 3.1%, 5.2%, and 8.4%. The proportion of patients with scores 0, 1, 2 and 3 and above in the 2003–2011 and 2012–2020 periods was 21.8% vs 28.5% (score 0), 41.6% vs 38.6% (score 1), 24.8% vs 23.1% (score 2) and 11.8% vs 9.8% (score ≥3), respectively (P=0.01).

Table 5.

Model score with respective Day-100 TRM rate

Score	Proportion of patients	Day-100 TRM rate
0	25.1%	0.46%
1	40.1%	3.2%
2	24%	5.8%
3 or greater	10.8%	14.5%

TRM, Transplant-related death

Comparison of the Day-100 TRM model with the cardiac stage model

We compared the prediction of the Day-100 TRM model with the European modification of the Mayo 2004 cardiac stage. The analysis includes 911 patients with available pre-ASCT troponin T or high-sensitivity troponin T (53% of the study cohort). Day-100 TRM model scores 0, 1, 2 and ≥3 was present in 30.7%, 34.6%, 23.6% and 11.1% of the patients respectively. In contrast, cardiac stages 1, 2, 3A and 3B were present in 37.9%, 45.2%, 12.5% and 4.4% of patients, respectively. The Day-100 TRM rates in this subanalysis were 0%, 3.5%, 6.5% and 19.8% for our model’s scores, 0, 1, 2, and ≥3, respectively. In comparison, the Day-100 TRM rates were 0.9%, 6.3%, 7% and 20% for cardiac stage 1, 2,3A and 3B, respectively. The AUC for the TRM model for predicting Day-100 TRM in this cohort subset was 0.79 compared to 0.71 for the Cardiac stage model (P=0.028).

Discussion

This study, the largest in the setting of ASCT in AL amyloidosis, was designed for predicting Day-100 TRM and is simple and easy to apply in routine practice. We identified four pre-ASCT predictors (NT-proBNP/BNP, ECOG PS, serum albumin, and systolic blood pressure), which, when grouped into one model predict Day-100 TRM with an AUC of 0.70.

The Day-100 TRM risk has significantly declined over the past three decades owing to improvements in patient selection and changes in clinical practice (i.e., induction before ASCT) and the availability of novel therapies with fewer toxicities. Improvement in supportive care in high-volume amyloidosis transplantation centers, led to a further decline in TRM. Organ failure as the cause of Day-100 TRM notably declined, supporting better patient selection and supportive care in the efforts to mitigate TRM. While exclusion criteria for ASCT emerge as safeguards to reduce TRM, they do not allow careful individualized estimation of the TRM risk. The reported TRM rate by several centers (1.1–5.0%)^{3, 7, 10, 30} is average and does not consider patient individual risk. Therefore, the Day-100 TRM model proposed in this work allows individualized risk assessment of early death following ASCT. This is relevant for clinical decision-making and allows for refinement in patient selection for safer transplantation. Acknowledging the benefits and risks of ASCT compared with alternative therapies will improve decision-making and potentially reduce further the TRM. With this new tool, patients can participate in shared decision-making by being accurately apprised of the mortality risks associated with proposed therapies. This is particularly relevant with the advent of newer therapies for AL amyloidosis, making the choice of ASCT one of several options. Previously, ASCT was chosen due to a lack of alternatives with a higher TRM in earlier years. As treatments for AL amyloidosis evolve and include plasma cell-directed therapies as well as anti-amyloid fibrils therapies, this tool should enhance informed decision-making on the role of ASCT.

The four predictors included in the model highlight the heterogeneous presentations of AL amyloidosis. These predictors are associated with the main causes of early death after ASCT: shock, high-grade arrhythmia, and organ failure. Blood pressure tends to decrease immediately post-ASCT,³¹ therefore pre-ASCT hypotension is expected to worsen during the post-ASCT course and predispose patients to a prolonged hypotension with tissue hypoperfusion. Cardiac AL amyloidosis (represented in this model by elevated NT-proBNP/BNP) is associated with high-grade arrhythmias and end-stage heart failure, both of which may be precipitated or worsened in the setting of high-dose chemotherapy.³² Heart involvement often leads to hypotension due to restrictive heart physiology with reduced compensatory mechanisms in the setting of pre-load or afterload reductions, which frequently develop in transplant patients.³³ Hypoalbuminemia is a novel risk factor for TRM identified in this study. It was present in nearly 20% of the study population. Hypoalbuminemia, which in this study was exclusively due to renal amyloidosis, leads to low oncotic pressure resulting in hypotension.^{34, 35} Hypoalbuminemia was previously identified as a risk factor for early progression to dialysis immediately after ASCT, thus associated with acute renal failure in the peritransplant period previously associated with high mortality.²¹ We have shown that ECOG PS 2/3 in this study was associated with heart involvement, advanced renal stage, autonomic nerve involvement, and low SBP. In the setting of low SBP, reduced tissue perfusion may play a role in the development of high-grade cardiac arrhythmias^{36, 37} and organ failure, making low SBP a contributor to the cause of death. Nonetheless, with the limitation of a retrospective determination of the cause of death, other factors play a role in the various causes of early death after ASCT, limiting the predictability of Day-100 TRM, as evident with an AUC of 0.7.

Attention to hypotension before and during ASCT is important in patients with AL amyloidosis and attempts to mitigate this risk should be considered. This includes avoidance of dehydration, monitoring of body weight, caution using anti-hypertensives, use of compression stockings and abdominal binder, and pharmacological management of autonomic neuropathy with midodrine, droxidopa, or fludrocortisone. The prevention of high-grade arrhythmias can include consideration of ICD placement prior to ASCT in selected patients at risk for high-grade arrhythmias. Continuous cardiac rhythm monitoring should be offered to those who are at risk of high-grade arrhythmias post-ASCT with immediate correction of precipitating factors such as anemia, electrolyte imbalance, and abrupt medication cessation.

Heart involvement is considered the dominant factor for short- and long-term survival. In our model, heart involvement is represented by the pre-ASCT natriuretic peptide level and was the second most influential factor in the TRM model. Other heart-related factors, such as EF and IVS, were potential candidates but were not included in the final model for simplicity and to minimize overlap. Of the two widely available cardiac blood biomarkers, troponin, and natriuretic peptide, troponin is used as an eligibility criterion for ASCT in AL amyloidosis, based on prior work showing early death discrimination.¹¹ However, due to the inability to reconcile the various troponin assays across institutions and periods, troponin was excluded from the modeling, limiting our understanding of its role in concert with other predictors, particularly NT-proBNP/BNP. A subset comparison of the TRM prediction model with the cardiac stage model in predicting Day-100 TRM demonstrated that the TRM model performs better, especially in patients with intermediate risk of TRM, such as those with a score of 1 or 2 in the current model and patients with cardiac stage 2 and 3A. In addition, a higher proportion of patients are included in the intermediate risk categories in the TRM model than in the cardiac staging model, further supporting its superior utility over the cardiac staging. This highlights that a model that encompasses the heterogeneity of AL amyloidosis does better in predicting outcomes than a model solely based on cardiac biomarkers.

A dFLC ≥180 mg/L at ASCT was associated with a higher rate of Day-100 TRM. We decided not to include it in our final model, recognizing the introduction of daratumumab has reduced the predictive value of dFLC on survival.^{28, 29} In today’s practice, with increasing treatment options, most patients who receive ASCT do so after prior therapy exposure and proceed to ASCT either due to lack of response to previous therapy or, more commonly, due to inadequate response depth following standard therapy. This makes the likelihood of having a pre-ASCT dFLC ≥180 mg/L less than in prior decades,³⁸. Therefore, we believe it is justified not to include dFLC in the model since a high proportion of patients will receive induction with a very high proportion achieving dFLC<180mg/dL prior to consideration of SCT.

Limitations of this work include its retrospective design, which may bias data collection and augment data missingness, for which imputation was used in some of the tested variables. The model also lacks external validation. External validation sources are difficult to obtain in rare diseases for whom only a small proportion of patients undergo ASCT. The validation of this model in the current era, where ASCT may be used later in the disease course, is uncertain and warrants future collaborative study.

In conclusion, we provide a simple clinical tool to predict the Day-100 TRM risk in AL patients undergoing ASCT. This tool will allow more informed decision making weighing the individual benefits and risks of ASCT compared to other therapies. This study also aids in awareness of the causes of early death after ASCT, which will allow the planning of more effective ways to prevent early deaths, namely hypotension, high-grade arrhythmias, and organ failure. We identified hypoalbuminemia as a novel risk factor for Day-100 TRM in AL amyloidosis, which is an important factor as many patients undergoing ASCT are patients with renal involvement in whom -generated hypoalbuminemia is common.

Data sharing statement:

The dataset of this study is not publicly available due to the proprietary nature of the electronic health records and maintaining patient anonymity mandated by the institutional review boards.