Medication use and risk of amyotrophic lateral sclerosis: using machine learning for an exposome-wide screen of a large clinical database

Abstract

Background

Accumulating evidence suggests that non-genetic factors have important etiologic roles in amyotrophic lateral sclerosis (ALS), yet identification of specific culprit factors has been challenging. Many medications target biological pathways implicated in ALS pathogenesis, and screening large pharmacologic datasets for signals could greatly accelerate the identification of risk-modulating pharmacologic factors for ALS.

Method

We conducted a high-dimensional screening of patients’ history of medication use and ALS risk using an advanced machine learning approach based on gradient-boosted decision trees coupled with Bayesian model optimization and repeated data sampling. Clinical and medication dispensing data were obtained from a large Israeli health fund for 501 ALS cases and 4,998 matched controls using a lag period of 3 or 5 years prior to ALS diagnosis for ascertaining medication exposure.

Results

Of over 1,000 different medication classes, we identified 8 classes that were consistently associated with increased ALS risk across independently trained models, where most are indicated for control of symptoms implicated in ALS. Some suggestive protective effects were also observed, notably for vitamin E.

Discussion

Our results indicate that use of certain medications well before the typically recognized prodromal period was associated with ALS risk. This could result because these medications increase ALS risk or could indicate that ALS symptoms can manifest well before suggested prodromal periods. The results also provide further evidence that vitamin E may be a protective factor for ALS. Targeted studies should be performed to elucidate the possible pathophysiological mechanisms while providing insights for therapeutics design.

Background

While substantial effort has been made to decipher the genetics of ALS, it is estimated that only about 10% of sporadic cases are due to monogenic mutations, while the etiology of the remainder remains unclear.(1–4) Different lines of evidence, including links to certain occupational exposures and very high but decreasing ALS rates in the western pacific(5–7), have led to the hypothesis that exogenous factors may contribute to ALS risk.(2, 8) However, the identification of non-genetic risk factors for ALS has been challenging.

The challenge of identifying risk and protective factors for ALS could be greatly accelerated by screening very large datasets for associations worthy of follow-up studies—effectively human equivalents of high throughput in vitro screening. An example of such high dimensional data is medication use, given that some medications may target biological pathways that are relevant for ALS pathogenesis. In vitro screening approaches of selected drug compounds with possible biological relevancy to ALS risk have thus far failed to translate findings into therapeutics.(9, 10) Few epidemiological studies have also explored the role of medications in ALS risk, but most focused only on a small subset of medications analyzed one-at-a-time,(11–15) which ignores correlation and interactions between medications.(16–18)

Therefore, we conducted a high-throughput screen of medications to elucidate possible individual and joint effects of medications on ALS onset. To do this, we adapted a gradient-boosted decision tree machine learning approach - eXtreme Gradient Boosting (XGBoost) - to analyze a large medication dispensing database nested within an electronic medical records system. To date, no similar screening approach has been undertaken to explore whether the vast array of commonly used medications, or combinations of such medications, is related to ALS risk.

Methods

Study design

Maccabi Health Services (MHS) is Israel’s second largest integrated health fund that provides universal medical services to 2.7 million members. An electronic medical record (EMR) has been implemented in MHS since the mid-90’s and contains detailed individual-level clinical and demographic information.

We defined patients with ALS as those meeting one of the following criteria: 1) had at least one ALS diagnosis record based on the International Classification of Disease, 9th Edition (ICD-9) diagnosis code 335.20 recorded by a neurologist or a physician specialist within a hospital; 2) had an ALS diagnosis and at least one dispensing record for riluzole, regardless of the diagnosing physician; or 3) had at least two diagnosis records for ALS recorded at different times regardless of the diagnosing physician. We used the above criteria to ascertain all ALS cases diagnosed between 2002–2020 who had at least four years of continuous membership in MHS prior to their date of diagnosis and complete demographic information, for a total of 501 cases. Previous studies demonstrated high validity of using physician diagnoses obtained from medical records to ascertain ALS cases compared with a review of medical records using the El Escorial criteria, considered as the gold-standard.(19, 20) The first date of ALS diagnosis or riluzole purchase was defined as the index date. We used incidence density matching to match 10 control patients to each ALS case. Controls had to be ALS free and have at least 4 years of continuous MHS membership prior to the index date. To isolate a possible effect of medication and improve computational efficiency, matching was based on the following criteria: birth year, sex, socioeconomic status derived from a poverty index based on each patient’s place of residence,(21) residential district, population subgroup (secular Israelis, Jewish orthodox, Israeli Arabs), and periphery index capturing the proximity of the residential location to large urban and economic centers.(22) The latter four demographic variables were included since the prevalence of reported ALS, age at recognition, and medication dispensing patterns are known to vary across these factors.(23, 24) For similar reasons, we also accounted for patients’ monthly average number of physician visits up to 3 years before the index date.(25) The study was approved by the institutional review boards of MHS and the Harvard Chan School of Public Health.

Medication information

We obtained the complete medication dispensing history for all study participants. Prescription medications are generally covered by the Israeli national health insurance and are offered free of charge or with minimal copays through more than 700 MHS-affiliated pharmacies. While over-the-counter (OTC) medications can be purchased at other locations, their cost is cheaper at MHS-affiliated pharmacies. To minimize the possibility that some medications may have been used because of early manifestation of ALS symptoms that preceded clinical diagnosis, we only included dispensing events recorded at least 3 years before the index date (a 5-year lag was used as a sensitivity analysis). To decrease the dimensionality of the data while preserving biological interpretability, we grouped the medications based on their active chemical substance according to their Anatomical Therapeutic Chemical (ATC) group.(26) To reduce noise in the data, we excluded 269 ATC groups that were used by less than 0.1% of the study participants, as these rarely used groups would not provide meaningful information to the analytic approach.

Statistical analysis

Conventional statistical approaches are of limited use with very high dimensional data. The simultaneous evaluation of many exposures requires that enough cases be observed for each exposure pattern, a condition that is seldom met for a rare outcome like ALS. This often results in unstable statistical models with limitedly interpretability.(17, 18) Machine learning approaches are specifically designed to overcome these common limitations arising with high-dimensional data. Thus, to conduct a screening of the medication data for possible ALS links, we used eXtreme Gradient Boosting (XGBoost), a highly efficient implementation of gradient tree boosting.(27) Briefly, a classification tree is a model that specifies repeated binary splits of predictors, assigning the same prediction to all observations (e.g. persons in the data) that end up in the same terminal node. Splitting is done across all predictors to optimize the model’s predictive accuracy, which is assessed through a prespecified loss function. Gradient boosting is an ensemble approach which combines the outputs of many individual trees (‘weak learners’) to capture complex nonlinearities and interactions. The combining of weak learners is done iteratively, so that each new tree is fit on the residuals of the previous one. XGBoost is a specific implementation of gradient boosting which uses advanced regularization and parallel processing to improve predictions and computing time. The model does not generally require special data preprocessing, but does require the specification of several data-specific regularization parameters to optimize performance.(28) We note that machine learning algorithms like XGBoost are typically used for predictive modeling.((29), also see additional discussion in e-materials concerning the difficulty of designing a prediction model for ALS using EMR data). However, our goal here was to optimize the use of the prediction framework to conduct a high-throughput screening for medications (individual or in combinations) linked with ALS as candidates for further research.

To reduce the likelihood of spurious findings, our analytic approach involved repeated sampling of the data using a nested cross validation (CV) design coupled with Bayesian model optimization (see e-materials for additional information). This resulted in a set of 100 XGBoost models that were independently trained on different data subsets. For each of these models, we identified medications that were predictive of the outcome based on the mean of absolute shapley additive explanations (SHAP) values, which quantify, for each person in the data, how much a given medication group contributed to the model’s risk prediction output (see e-materials for additional information).(30) We extracted medication groups that were found to be associated with the outcome in at least 50% of the independently run models to focus on predictors consistently associated with the outcome across the independent runs, and therefore less likely to have occurred by chance.

While focusing on consistently associated medication classes better ensures reproducibility, a potential downside is the possibility of missing rarely used medications even if they are associated with the outcome, since users of these medications may be absent in most data splits. Similarly, commonly used medications may also be missed if the association with the outcome is primarily driven by a small subset of users with unique characteristics (e.g. genetic background). To mitigate this, we also selected medication groups deemed very important for outcome prediction in at least one run, defined based on having a mean absolute SHAP value in the upper 1% of the run-specific distribution after excluding groups with zero values (see e-materials). Once a final list of important medication groups was derived, an XGBoost model with these groups only was trained on the full data using the same optimization approach to derive the optimal value for the tree depth hyperparameter, which indicates the order (e.g. 2-way, 3-way, etc.) of interactions between the medication groups in the model. Since machine learning algorithms like XGBoost are not done in a hypothesis framework and do not provide specific effect estimates for each variable as in classical regression, the final list of predictors (including interactions if indicated) was then used in a conditional logistic regression model to estimate odds ratios and 95% confidence intervals for the association with ALS.

Results

Characteristics of the study population are presented in Table 1. The average age at first ALS diagnosis was 61.5, with more male ALS cases. The average number of follow-up months and frequency of physician visits were overall similar for cases and controls, while cases had a slightly higher number of dispensed ATC groups. In total, study participants had used 10,210 different medications, classified into 1,053 ATC groups. Of these, only 8 groups were consistently associated with ALS across the independent runs (i.e. in >50% of runs). An additional 10 medication groups not consistently associated with ALS had SHAP values in the top 1% of the distribution in at least one run and were thus included in the final list of medications (Table 2). All 8 consistently selected medications were more commonly used in cases compared to controls, with use prevalence in cases ranging from 1.4 to 30.3%. These 8 medication groups were also consistently selected when using a 5-year lag to ascertain medication exposure. Intriguingly, one of these 8 medications was Interferon beta-1a, a treatment for multiple sclerosis (MS), which is sometimes an initial (incorrect) diagnosis in someone who later receives an ALS diagnosis. Exclusion of ALS cases and controls who ever received an MS diagnosis resulted in two medications, interferon beta-1a and sumatriptan, dropping out of the list of associated medications.

Table 1.

Population characteristics of ALS cases and matched controls

	Cases (n= 501)	Controls (n=4998)
Sex, n (%)
Men	294 (58.7)	2934 (58.7)
Women	207 (41.3)	2064 (41.3)
Population subgroup, n (%)
Secular Jewish	458 (91.4)	4569 (91.4)
Jewish orthodox	28 (5.6)	279 (5.6)
Israeli Arabs	15 (3.0)	150 (3.0)
Periphery residence, n (%)	27 (5.4)	259 (5.2)
District of residence, n (%)
Center	125 (25.0)	1244 (24.9)
Jerusalem	114 (22.8)	1140 (22.8)
North	82 (16.4)	819 (16.4)
Sharon	116 (23.2)	1155 (23.1)
South	64 (12.8)	640 (12.8)
Socioeconomic Status, mean (sd)	6.6 (2.0)	6.6(1.9)
Age at index date, mean (sd)	61.5 (14.1)	61.5 (14.0)
Months of continuous follow-up prior to index date, mean (sd)	103 (40.4)	102 (40.8)
Average number of physician visits per month over follow-up perioda, mean (sd)	0.7 (0.6)	0.6 (0.5)
Number of unique medications dispensed per persona,b, mean (sd)	68.6 (62.1)	60.4 (55.6)
Number of unique ATC groups dispensed per persona,b, mean (sd)	43.2 (31.6)	38.5 (29.1)

^a.Excluding the 3 years immediately preceding the index date

^b.Through the beginning of follow-up, including over the counter (OTC) medications.

Table 2.

ATC groups consistently associated with the outcome based on global SHAP values

ATC group code	Main anatomic / pharmacologic group	Pharmacologic / therapeutic subgroup	chemical substance	Prevalence of use in cases (n, (%))	Prevalence of use in controls (n, (%))	% of runs in which association with ALS was observeda	% of runs with global SHAP value ≥99% of the distributionb
m01ah02	Musculo-skeletal system	Anti-inflammatory and Antirheumatic Products, Non-Steroids	Rofecoxib	152 (30.3%)	1050 (21.0%)	96	75
m03bx01	Musculo-skeletal system	muscle relaxants	Baclofen	17 (3.4%)	31 (0.6%)	82	0
a06ad10	Alimentary tract and metabolism	drugs for constipation	osmotically acting laxatives	65 (13.0%)	362 (7.2%)	75	7
n02cc01	Nervous system	Antimigraines	Sumatriptan	18 (3.6%)	55 (1.1%)	73	0
l03ab07	Immunomodulating agents	Immuno-stimulants	Interferon beta-1a	7 (1.4%)	3 (0.1%)	70	0
r05da20	Respiratory system	Cough suppressants	opium alkaloids and derivatives	117 (23.4%)	847 (16.9%)	65	11
a11db	Alimentary tract and metabolism	B vitamins	vitamin b1 in combination with vitamin b6 and/or vitamin b12	112 (22.4%)	779 (15.6%)	64	3
j01ee01	Anti-infective for systemic use	Sulfonamides and trimethoprim	sulfamethoxazole and trimethoprim	89 (17.8%)	614 (12.3%)	51	1
d07ca01	Dermatologicals	Corticosteroids with antibiotics	Hydrocortisone and antibiotics	82 (16.4%)	566 (11.3%)	42	1
a10ba02	Alimentary tract and metabolism	Blood glucose lowering drugs	Metformin	53 (10.6%)	661 (13.2%)	28	2
a11ha03	Alimentary tract and metabolism	Vitamins	Tocopherol (vit E)	58 (11.6%)	622 (12.4%)	20	1
a10bb01	alimentary tract and metabolism	Blood glucose lowering drugs	Glibenclamide	18 (3.6%)	278 (5.6%)	19	1
r01ba53	Respiratory system	Systemic nasal decongesants	Phenylephrine	213 (42.5%)	1835 (36.7%)	17	1
c10aa05	Cardiovascular system	Lipid modifying agents	Atorvastatin	69 (13.8%)	724 (14.5%)	15	1
d07cb01	Dermatologicals	Corticosteroids with antibiotics	Triamcinolone and antibotics	80 (16.0%)	807 (16.1%)	15	1
r05cb02	Respiratory system	Expectorants	Bromhexine	80 (16.0%)	788 (15.8%)	11	1
d02ab	Dermatologicals	Emollients and protectives	Zinc products	34 (6.8%)	343 (6.9%)	10	1
m01ae01	Musculo-skeletal system	Anti-inflammatory and Antirheumatic Products, Non-Steroids	Ibuprofen	189 (37.7%)	1769 (35.4%)	10	1

^a.Based on global SHAP values calculated for each of the 100 runs. Global SHAP value > 0 for a given feature was considered as evidence for an association with the outcome.

^b.Considering each of the 100 different global SHAP values distributions individually

After omitting the two possible MS-related medications, the correlations (Phi-coefficients) between the remaining 16 selected medication groups were low to moderate, ranging from 0.01 to 0.51. Bayesian optimization of an XGBoost model trained with these 16 medication groups suggested an optimal model with a maximum tree depth of 1, indicating no meaningful predictive interactions. Results from a mutually adjusted conditional logistic regression model for the 16 medication groups indicated elevated effect estimates for all 8 consistently selected medication groups that ranged from 1.3 to 5.1 (Table 3). Additionally, two of the inconsistently selected medication groups included in the final model also met the conventional threshold for statistical significance. Among them was Tocopherol (vit E), with a suggestive protective effect.

Table 3.

Odds Ratios^a (OR) and 95% confidence intervals (CI) from conditional logistic regression for selected ATC groups.

ATC group code	Main anatomic / pharmacologic group	Pharmacologic / therapeutic subgroup	chemical substance	OR (95% CI)	P-value
m01ah02	Musculo-skeletal system	Anti-inflammatory and Antirheumatic Products, Non-Steroids	Rofecoxib	1.51 (1.20–1.91)	<0.01
m03bx01	Musculo-skeletal system	muscle relaxants	Baclofen	5.07 (2.65–9.68)	<0.01
a06ad10	Alimentary tract and metabolism	drugs for constipation	osmotically acting laxatives	1.98 (1.44–2.73)	<0.01
r05da20	Respiratory system	Cough suppressants	opium alkaloids and derivatives	1.32 (1.04–1.69)	0.02
a11db	Alimentary tract and metabolism	B vitamins	vitamin b1 in combination with vitamin b6 and/or vitamin b12	1.47 (1.14–1.89)	<0.01
j01ee01	Anti-infective for systemic use	Sulfonamides and trimethoprim	sulfamethoxazole and trimethoprim	1.33 (1.02–1.74)	0.04
d07ca01	Dermatologicals	Corticosteroids with antibiotics	Hydrocortisone and antibiotics	1.37 (1.04–1.79)	0.02
a10ba02	Alimentary tract and metabolism	Blood glucose lowering drugs	Metformin	0.79 (0.56–1.13)	0.20
a11ha03	Alimentary tract and metabolism	Vitamins	Tocopherol (vit E)	0.65 (0.47–0.90)	0.01
a10bb01	alimentary tract and metabolism	Blood glucose lowering drugs	Glibenclamide	0.68 (0.39–1.19)	0.18
r01ba53	Respiratory system	Systemic nasal decongesants	Phenylephrine	1.11 (0.90–1.37)	0.32
c10aa05	Cardiovascular system	Lipid modifying agents	Atorvastatin	0.82 (0.61–1.11)	0.20
d07cb01	Dermatologicals	Corticosteroids with antibiotics	Triamcinolone and antibotics	0.77 (0.58–1.01)	0.06
r05cb02	Respiratory system	Expectorants	Bromhexine	0.82 (0.62–1.08)	0.16
d02ab	Dermatologicals	Emollients and protectives	Zinc products	0.82 (0.56–1.21)	0.32
m01ae01	Musculo-skeletal system	Anti-inflammatory and Antirheumatic Products, Non-Steroids	ibuprofen	0.91 (0.74–1.13)	0.39

^a.Controlled for birth year (+/− 2 years), sex, socioeconomic status (+/− 1 scale points) residential district, population subgroup and periphery index by matching, with additional adjustment for birth year and socioeconomic status to account for residual confounding due to matching on a range of these variables, and for the average number of monthly physician visits over the duration of the follow-up period up to 3 years prior to index date using a natural spline with 4 degrees of freedom.

Discussion

Using gradient boosting techniques and repeated data sampling, we identified several medication classes that were associated with ALS. Notably, an increased risk was observed with several medications that are clinically indicated for management of early non-specific symptoms that could be related to ALS. These includes rofecoxib, a selective COX-2 inhibitor that was taken off the market in 2005; baclofen, a muscle relaxant used to control muscle cramps, spasms, and spasticity; and osmotically acting laxatives used for constipation. Increased risks were also observed for Thiamine (vitamin B1) formulations, which are sometimes prescribed to patients complaining of neuropathy symptoms that are relevant for ALS.(31, 32) We also observed increased risks for some dermatologicals. Skin changes have been reported in ALS, possibly due to shared embryonic origin from the ectodermal germ layer, and may precede onset of neurological symptoms.(33–37)

We observed several medication classes that had protective associations with ALS, although we note that these medications were not consistently identified across the independently trained models. While the inconsistency increases chances of spurious findings, it can also occur for medications that are truly associated with ALS but are absent from most data partitions due to rare use, or for medications whose association with ALS is driven by some small subset of individuals with specific characteristics—that is, modifiers of any effect of the medication. Among these inconsistently selected medications, the strongest statistical evidence was for Vitamin E. Vitamin E has important antioxidant properties, and a role for oxidative stress in ALS pathogenesis has been suggested.(38) Some epidemiologic studies have linked high vitamin E intake with lower ALS risk,(39) and vitamin E supplementation was also associated with lower functional deterioration rate in randomized clinical trials of ALS patients.(40, 41) We also observed suggestive protective effects for metformin, an antidiabetic drug that was recently associated with improvement of ALS symptoms in mice,(42) and for atorvastatin. Statin use has been linked with protective effects on ALS incidence and survival, although not in all studies.(11, 43–46)

Our observations of increased ALS risk for medications that are clinically indicated for management of early ALS symptoms were consistent when using 3- and 5-year lag periods. Previous reports have suggested a median time of approximately 12 months from first ALS symptom to diagnosis.(47) Even if we allow for some additional diagnostic delay in our data, it is very unlikely that this would be enough to advance onset to 3- or 5-years before clinical recognition. Furthermore, the highly accessible nature of the Israeli medical system decreases the likelihood of missing or delaying a diagnosis. Estimates of a pre-clinical phase of ALS suggest some biomarker changes at most 1–2 years before onset.(48–50) If our findings result from underlying changes related to very early unrecognized ALS rather than mechanisms that alter ALS risk, then they suggest that such changes occur well before current evidence of pre-symptomatic biological alterations.

An intriguing finding was the increased ALS risk among users of sumatriptan and Interferon beta-1a. Sumatriptan is used to treat cluster headaches and migraines. Headaches are sometimes implicated in later ALS stages due to respiratory insufficiency, but cluster headaches are more consistently linked with early multiple sclerosis (MS) symptoms,(51, 52) for which Interferon beta-1a is an immunomodulator treatment. It is possible that use of these medications could relate to suspicion of MS, for which there can be clinical ambiguity with ALS in initial stages.(53) Supporting this hypothesis, neither of these medications were identified when individuals who ever received an MS diagnosis were excluded.

Overall, our results are in partial agreement with a previous report that screened Medicare prescription drugs, which reported 10 drugs that were associated with ALS risk using a 1-year and 3-year lag analysis.(15) These included drugs for hypertension and cardiovascular disease that were not identified in our analysis. In contrast, the Medicare analysis linked baclofen with increased ALS risk, and observed protective associations for metformin and simvastatin, similar to our results.(15) The partial discrepancy in the results between the analyses could be due to differences in the study population, prescription practices, recruitment period, and analytic approach. The Medicare analysis ascertained medication use over a median period of 49 months compared with 104 months in our analysis, and was restricted to ALS cases diagnosed in 2008–2014 at ages 66–89. Moreover, the Medicare analysis examined each medication separately using a series of independent regression models. This may lead to more medications being detected, as the approach does not account for potential correlation and confounding by other medications, a problem that our approach helps to avoid. Finally, even with multiple testing corrections, spurious findings could arise due to model overfitting when the data are only analyzed once as in the prior study. Our approach used cross-validation with repeated data sampling, a powerful technique to protect against overfitting and spurious findings.(54, 55)

Our analysis has several limitations. While we used extensive matching for confounding control, confounding bias is still possible. That said, the employed nested cross-validation approach serially resampled subsets of the data. Any association driven by confounding would likely be less consistently observed because it would be sensitive to the distribution of any confounding variables among the selected subset of cases and controls in each run—a causal effect of a medication would not vary by the distribution of other variables in each subsample. Therefore, our primary focus on consistently identified medication classes offers some protection against this issue. Confounding by indication, however, remains possible. Use of a medication is very strongly correlated with the condition for which the medication is prescribed. Therefore, additional targeted studies are needed to replicate the findings in other cohorts and decipher the underlying mechanisms driving the results.

Additional limitations include, first, that we considered any history of medication use, without restricting to specific time periods. If ALS is only associated with medications use during specific time periods, considering all historical dispensing data may attenuate any associations. Relatedly, we did not have information on dispensing events occurring prior to 1998, and thus did not capture early life exposures. Second, despite a financial incentive to purchase OTC medications through MHS pharmacies, we cannot guarantee complete ascertainment for these medications. Third, we considered use of medications as a dichotomized contrast without considering duration of use or dose, factors that may also affect ALS risk. All of these could have introduced some error in medication exposure assignment, but this is likely to be the same for cases and controls and thus, if anything, likely to only attenuate any true associations.(56)

In conclusion, using a high-dimensional screening of dispensed medications, we found evidence for an increased ALS risk with several medications even well before the likely onset of symptoms. This suggests either that early signs of ALS can manifest years before the condition is clinically recognized, or that these medications are worth exploring further for their potential role as ALS risk factors. We also found suggestive protective associations, particularly with vitamin E. Our findings not only have implications for ALS pathogenesis, but also show that the described gradient boosting approach could be a powerful tool for exploring multidimensional data for exposure-disease relationships to advance etiological understanding and the development of possible therapeutics.