A NOVEL AFROCENTRIC STROKE RISK ASSESSMENT SCORE: MODELS FROM THE SIREN STUDY

Abstract

Background:

Stroke risk can be quantified using risk factors whose effect sizes vary by geography and race. No stroke risk assessment tool exists to estimate aggregate stroke risk for indigenous African.

Objective:

To develop Afrocentric risk-scoring models for stroke occurrence.

Methods:

We evaluated 3,533 radiologically confirmed West African stroke cases paired 1:1 with age-, and sex-matched stroke-free controls in the SIREN study. The 7,066 subjects were randomly split into a training and testing set at the ratio of 85:15. Conditional logistic regression models were constructed by including 17 putative factors linked to stroke occurrence using the training set. Significant risk factors were assigned constant and standardized statistical weights based on regression coefficients (β) to develop an additive risk scoring system on a scale of 0 to 100%. Using the testing set, Receiver Operating Characteristics (ROC) curves were constructed to obtain a total score to serve as cut-off to discriminate between cases and controls. We calculated sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) at this cut-off.

Results:

For stroke occurrence, we identified 15 traditional vascular factors. Cohen’s kappa for validity was maximal at a total risk score of 56% using both statistical weighting approaches to risk quantification and in both datasets. The risk score had a predictive accuracy of 76% (95%CI: 74 – 79%), sensitivity of 80.3%, specificity of 63.0%, PPV of 68.5% and NPV of 76.2% in the test dataset. For ischemic strokes, 12 risk factors had predictive accuracy of 78% (95%CI: 74–81%). For hemorrhagic strokes, 7 factors had a predictive accuracy of 79% (95%CI: 73–84%).

Conclusion:

The SIREN models quantify aggregate stroke risk in indigenous West Africans with good accuracy. Prospective studies are needed to validate this instrument for stroke prevention.

1.0. INTRODUCTION

The occurrence of stroke is driven by a combination of socio-demographic and cardio-metabolic factors whose contributions vary by geographical region and by race. Typically, clustering of risk factors in an individual is directly related to the probability of developing a stroke. There are several cardiovascular and stroke risk prediction models including the Framingham¹ (USA), Prospective Cardiovascular Munster (PROCAM) study (USA)², Global vascular risk score (GVRS) from the Northern Manhattan Study (USA)³, Reynolds risk score (USA)⁴, the China Multicenter Collaborative Study of Cardiovascular Epidemiology (MUCA) (China)⁵, the ASSIGN score from Scotland⁶, CUORE from Italy⁷, QRISK from England and Wales⁸, SCORE from Europe⁹ and finally a score from a Korean national health examination data.¹⁰ All these risk calculators were developed using prospectively collected data by using regression coefficients and relative risks for significant risk factors in Cox proportional models.

Conspicuously missing from the literature is a stroke prediction model for indigenous Africans.¹¹ Stroke in sub-Saharan Africa has several distinctive features including earlier age of onset, proclivity towards severe strokes and high short-term mortality.¹² Inadequate neurologists, paucity of acute stroke services, rehabilitation and social services means that a strategy to detect those at risk in the population for primary prevention is an urgent operational research priority. However, the perennial lack of resources to conduct prospective population-based studies in Africa means stroke prediction models should first begin from case-control studies for later validation in prospective cohorts. With this in mind, we sought to develop risk-scoring models for quantifying individuals’ aggregate risk of having a stroke and its two primary types.

2.0. METHODS

2.1. Study Design

The Stroke Investigative Research and Educational Network (SIREN) study is a case-control study involving 15 centers in Ghana and Nigeria.¹³ Ethical approval was obtained from all study sites and informed consent was obtained from all participants. Stroke cases were adults ≥18 years with first clinical stroke, with neuroimaging confirmation using CT or MRI scan within 10 days of symptom onset. Controls were consenting stroke-free adults who were recruited mostly from the communities in the catchment areas of the SIREN hospitals where stroke cases resided. Stroke-free status was ascertained using a locally pre-validated 8-item questionnaire for verifying stroke-free status (QVSFS).¹⁴ We matched stroke-free controls by age (±5 years), sex and ethnicity to minimize the potential confounding effect of these variables on the relationship between stroke and its risk factors.

2.2. Selection of vascular risk factors

The dataset on the 7,066 matched case-control subjects were randomly split into a training set (6,040) and testing set (1,066) at the ratio of 85:15. ¹⁵ The training set was used to identify putative stroke risk factors and to develop the risk quantification model while the testing set was used to test and evaluate the performance of method.

Seventeen (17) potential risk factors were initially tested for associations with stroke occurrence using the training dataset (Table 1). Procedures used to measure presence and definitions of these risk factors in our study have been previously described and briefly presented under the Supplementary information.¹³ These risk factors were derived from an extensive literature review, our clinical understanding of stroke risk, and empirical evidence (significant associations observed in bivariate analyses). Bivariate analyses on matched case-control pairs were conducted on the training dataset following the methods described by Greenberg and Ibrahim.¹⁶ A conditional multiple logistic regression analysis was then performed to assess which of the 17 putative factors were independently associated with stroke occurrence in the full model. Then a final model that included the risk factors attaining a significant α level of 0.05 was constructed (Table 1).

Table 1:

Conditional multiple logistic regression analysis for All stroke types using training data

Factors	aOR	95%CI	p-Value
Variables Reported in the Literature
Baseline age ≥50 y	2.27	(1.30–3.96)	0.004
Income level, <$100 USD	1.61	(1.32–1.98)	<0.001
Education (none vs some)	1.38	(1.04–1.82)	0.02
Hypertension	15.08	(10.56–21.53)	<0.001
Dyslipidemia	3.33	(2.63–4.21)	<0.001
Diabetes mellitus	2.82	(2.25–3.54)	<0.001
Cardiac disease	1.74	(1.21–2.52)	0.002
Family history of CVD	1.50	(1.20–1.87)	<0.001
Raised waist-hip ratio	1.71	(1.35–2.16)	<0.001
Current tobacco use	2.02	(0.98–4.14)	0.05
Current alcohol use	0.97	(0.77–1.22)	0.77
Stress	1.44	(1.12–1.84)	0.003
sprinkling salt on food at table	1.68	(1.13–2.50)	0.01
Low consumption of leafy green vegetable	2.67	(2.10–3.39)	<0.001
Regular sugar consumption	1.24	(1.01–1.53)	0.03
Meat consumption	1.75	(1.36–2.26)	<0.001
Physical inactivity	1.68	(1.02–2.76)	0.04
Final Model
Baseline age ≥50 y	2.18	(1.26–3.79)	0.005
Income level, <$100 USD	1.59	(1.30–1.95)	<0.001
Education (none vs some)	1.37	(1.04–1.82)	<0.001
Hypertension	15.47	(10.84–22.07)	<0.001
Dyslipidemia	3.37	(2.67–4.26)	<0.001
Diabetes mellitus	2.79	(2.22–3.50)	<0.001
Cardiac disease	1.70	(1.18–2.45)	0.004
Family history of CVD	1.50	(1.20–1.86)	<0.001
Raised waist-hip ratio	1.70	(1.34–2.15)	<0.001
Stress	1.45	(1.13–1.85)	0.003
sprinkling salt on food at table	1.71	(1.15–2.54)	0.008
Low consumption of leafy green vegetable	2.66	(2.10–3.38)	<0.001
Regular sugar consumption	1.25	(1.01–1.53)	0.03
Meat consumption	1.78	(1.38–2.29)	<0.001
Physical inactivity	1.71	(1.04–2.81)	0.03

aOR- Adjusted Odds Ratio; CI- Confidence Interva

2.3. Development of a scoring system for stroke risk factors

The significant risk factors in the final model were then given statistical weights derived for each factor by a linear transformation of the regression coefficient for that factor in the final conditional logistic regression model. This linear transformation was essential to enhance interpretability and make the scoring system easy to use. We compared two approaches to assigning statistical weights. The first was a constant weighting (β_c) obtained by rounding up β by multiplying by 10 and secondly by obtaining a standardized weighted β coefficient, ${\beta }_w=\beta \left(\frac{S_x}{S_y}\right)$ where β_w= standardized beta coefficient, β = estimated coefficient, S_x =Standard deviation of predictive variable and S_y=Standard deviation of response variable.

Each of the stroke cases and stroke-free controls in the training datasets were then scored individually using the risk scoring system developed. To enhance comparability and interpretability, scores were transformed on the scale of 0–100.¹⁷ To calculate the interval for the scores, we applied the Sturges rule.^18–19Thus, for both weighting approaches, the interval obtained was 7.3 ≈ 7. The predictive accuracies, sensitivities, specificities and Cohen’s kappa of the risk scoring system at various cut-offs of the aggregate score were calculated.²¹

In order to evaluate the performance of the scoring system, each of the 1,066 stroke cases and stroke-free controls in the test dataset were also scored individually using the risk scoring system (developed from the training data). The intervals obtained from the training set were applied to the scores obtained from the test set. The predictive accuracies, sensitivities, specificities and Cohen’s kappa of the risk scoring system were calculated for the same cut-offs used for the training set. The overall predictive accuracy was calculated as an equivalent of the Wilcoxon statistics as described by Hanley and McNeil.²² The best cut-off for the total score was obtained by plotting Receiver Operator Characteristics (ROC) Curves and the sensitivity, specificity, positive predictive accuracy and negative predictive accuracy were estimated at the cut-off points and presented for the training and test sets. Similar approaches were used for ischemic strokes and hemorrhagic strokes. All data analyses were performed using the R software (version 3.6.2) and IBM SPSS Statistics version

3.0. RESULTS

3.1. Risk scoring models for stroke occurrence

Fifteen (15) out of the 17 putative factors associated with stroke occurrence were included in our final model using the training dataset (Table 1). These were baseline age > 50 years, income level <$100, lack of formal education, hypertension, dyslipidemia, diabetes mellitus, cardiac disease, family history of cardiovascular disease, raised waist-hip ratio, stress, sprinkling salt on food at table, low-consumption of green leafy vegetable, regular sugar consumption, regular meat consumption and physical inactivity. Table 2 shows statistical weights assigned to each of the significant risk factors after linear transformation using either constant weighting or standardized weighting.

Table 2:

A scoring system for all stroke types using training data

Factors from the Final Model	β	Constant Weighting	Standardized Weighting (βw)
		(10*β)	βw	10*βw
Baseline age ≥50 y	0.78	8	0.68	7
Income level, <$100 USD	0.47	5	0.46	5
Education (none vs some)	0.32	3	0.25	3
Hypertension	2.74	27	2.31	23
Dyslipidemia	1.22	12	1.12	11
Diabetes mellitus	1.03	10	0.86	9
Cardiac disease	0.53	5	0.29	3
Family history of CVD	0.40	4	0.38	4
Raised waist-hip ratio	0.53	5	0.45	4
Stress	0.37	4	0.28	3
sprinkling salt on food at table	0.54	5	0.29	3
Low consumption of leafy green vegetable	0.98	10	0.81	8
Regular sugar consumption	0.22	2	0.21	2
Meat consumption	0.58	6	0.43	4
Physical inactivity	0.54	5	0.21	2

β - Regression coefficients; β_c – Constant-weighted Regression Coefficient; β_w – Standardized-weighted Regression Coefficient

The range of possible total raw scores for an individual by our risk scoring system was 0 to 104 or 0 to 86 using either the constant weighting or standardized weighting approach respectively (and 0 to 100 for the transformed score). It should be noted that all predictive variables were dichotomized as either present or absent for each subject. Table 3 shows the classification of stroke cases and controls by the total risk score categories using a total score of 7 as the class intervals for the weighting approaches.

Table 3:

Classification of study subjects by total risk score for all stroke types in Training and Test Datasets

		Training Dataset		Test Dataset
	Score Range	Control (N=3020)	Cases (N=3020)	Control (N=533)	Cases (N=533)
Constant weighting [10*β]	0 – 6	13	7	1	0
	7 – 13	107	23	13	1
	14 – 20	222	19	19	11
	21 – 27	304	21	36	2
	28 – 34	315	46	49	11
	35 – 41	301	64	54	12
	42 – 48	361	144	70	22
	49 – 55	414	307	68	43
	56 – 62	429	585	92	97
	63 – 69	360	882	45	100
	70 – 76	134	547	43	70
	77 – 83	53	301	38	88
	84 – 90	7	65	5	52
	91 – 97	0	8	0	21
	98 – 100	0	1	0	3
Standardized weighting [10*βw]	Score Range	Control (N=3020)	Cases (N=3020)	Control (N=533)	Cases (N=533)
	0 – 6	13	7	4	0
	7 – 13	90	22	10	1
	14 – 20	201	17	32	12
	21 – 27	332	19	29	1
	28 – 34	272	36	49	13
	35 – 41	294	63	50	9
	42 – 48	333	121	70	25
	49 – 55	397	237	92	44
	56 – 62	419	506	65	92
	63 – 69	361	719	48	92
	70 – 76	196	647	49	90
	77 – 83	88	459	27	73
	84 – 90	22	135	6	52
	91 – 97	2	30	2	27
	98 – 100	0	2	0	2

β_c – Constant-weighted Regression Coefficient; β_w – Standardized-weighted Regression Coefficient

The validity characteristics of the risk scoring system at various cut-offs are shown for both the training and test sets in Table 4. The Cohen’s kappa for validity was maximal at a total risk score of 56% for both the constant and the standardized weighting methods. At this cut-off, the scoring system yielded a sensitivity score of 80.9% (specificity =58.2%) for the constant weighting, while for the standardized weighting, sensitivity was 80.3% (specificity =63.0%) for the test dataset. The ROC curve for the scoring system using the two weighting approaches produced the same AUC of 0.76, p<0.001 for the test dataset (Table 4 and Figure 1).

Table 4:

Performance of the risk score for all stroke types using training and test dataset

		Training Dataset					Test Dataset
	Cut off for the risk score	Se%	Sp%	PPV %	NPV %	K	Se%	Sp%	PPV %	NPV %	K
Constant weighting [10*β]	≥ 7	99.8	0.4	50.0	65.0	0.002	100.0	0.2	50.0	100	0.002
	≥ 14	99.0	4.0	50.8	80.0	0.030	99.8	2.6	50.6	93.3	0.024
	≥ 21	98.4	11.3	52.6	87.5	0.097	97.7	6.2	51.0	73.3	0.039
	≥ 28	97.7	21.4	55.4	90.2	0.191	97.4	12.9	52.8	83.1	0.103
	≥ 35	96.2	31.8	58.5	89.2	0.280	95.3	22.1	55.0	82.5	0.174
	≥ 42	94.0	41.8	61.8	87.5	0.358	93.1	32.3	57.9	82.3	0.253
	≥ 49	89.3	53.7	65.9	83.4	0.430	88.9	45.4	62.0	80.4	0.343
	≥ 56	79.1	76.3	70.8	76.3	0.466	80.9	58.2	65.9	75.2	0.390
	≥63	59.7	81.7	76.5	67.0	0.414	62.7	75.4	71.8	66.9	0.381
	≥ 70	30.5	93.6	82.6	57.4	0.241	43.9	83.9	73.1	59.9	0.278
	≥ 77	12.4	98.0	86.2	52.8	0.104	30.8	91.9	79.2	57.0	0.227
	≥ 84	2.5	99.8	91.4	50.6	0.022	14.3	99.1	93.8	53.6	0.133
	≥ 91	0.3	100	100	50.1	0.003	4.5	100	100	51.2	0.045
	≥98	0.0	100	100	50.0	0.000	0.6	100	100	50.1	0.006
	AUC (95%CI): 0.80(0.79 –0.81); p<0.001						AUC (95%CI): 0.76(0.73– 0.79); p<0.001
Standardized weighting [10*βw]	Cut off for the risk score	Se%	Sp%	PPV %	NPV %	K	Se%	Sp%	PPV %	NPV %	K
	≥ 7	99.8	0.4	50.0	65.0	0.002	100	0.8	50.2	100	0.008
	≥ 14	99.0	78.0	50.6	78.0	0.025	99.8	2.6	50.6	93.3	0.024
	≥ 21	98.5	86.9	52.3	86.9	0.085	97.6	8.6	51.6	78.0	0.062
	≥ 28	97.8	21.1	55.3	90.7	0.189	97.4	14.1	53.1	84.3	0.114
	≥ 35	96.7	30.1	96.7	90.0	0.267	94.9	23.3	55.3	82.1	0.182
	≥ 42	94.6	39.8	61.1	88.0	0.344	93.2	32.6	58.1	82.9	0.259
	≥ 49	90.6	50.8	64.8	84.3	0.414	88.6	45.8	62.0	80.0	0.343
	≥ 56	82.7	64.0	69.7	78.7	0.467	80.3	63.0	68.5	76.2	0.433
	≥63	66.0	77.8	74.9	69.6	0.438	63.0	75.2	71.8	67.1	0.383
	≥ 70	42.2	60.8	80.5	60.8	0.320	45.8	84.2	74.4	60.8	0.300
	≥ 77	20.7	96.3	84.8	54.8	0.170	28.9	93.4	81.5	56.8	0.223
	≥ 84	5.5	99.2	87.4	51.2	0.047	15.2	98.5	91.0	53.7	0.137
	≥ 91	1.1	99.9	94.1	50.2	0.010	5.4	99.6	93.5	51.3	0.051
	≥98	0.1	100	100	50.0	0.001	0.4	100	100	50.1	0.004
	AUC (95%CI): 0.80(0.79 – 0.81); p<0.001						AUC (95%CI): 0.76(0.74–0.79); p<0.001

S_e-Sensitivity; S_p-Specificity; PPV-Positive Predictive Value; NPV-Negative Predictive Value; K-Cohen’s Kappa; AUC-Area Under the ROC; CI-Confidence Interval.

Fig. 1.

ROC Curves for training and testing dataset for all strokes using constant and standardised weighting.

3.2. Risk score model for occurrence of ischemic strokes

Among 2,203 ischemic strokes and matching controls, 12 risk factors were included in the final model using the training set (Supplementary Table 1). The statistical weights and classification of stroke cases and controls using the scoring systems are shown in Tables S2 and S3. The validity characteristics of the risk scoring system at various cut-off are presented for the training and test datasets in Table 5. The Cohen’s kappa was maximal at a total risk score of 56% for both the constant and the standardized weighting methods with AUCs of 0.78, p<0.001 (Figure 2) for the test dataset.

Table 5:

Performance of the risk score for Ischemic stroke types using training and test dataset

		Training Dataset						Test Dataset
	Cut off for the risk score	Se%	Sp%	PPV %	NPV %	K		Se%	Sp%	PPV %	NPV %	K
Constant weighting [10*β]	≥ 8	99.7	1.5	50.3	82.4	0.012		100	0.9	50.2	100	0.009
	≥ 16	99.2	6.4	51.5	88.9	0.056		98.5	3.6	50.5	70.6	0.021
	≥ 24	98.5	17.3	54.4	92.0	0.158		97.6	12.4	52.7	83.7	0.100
	≥ 32	97.3	24.9	56.5	90.3	0.223		97.0	18.8	54.4	86.1	0.158
	≥ 40	93.3	42.1	61.7	86.3	0.354		93.9	31.8	57.9	84.0	0.258
	≥ 48	89.2	54.3	66.1	83.4	0.435		90.6	47.3	63.2	83.4	0.379
	≥ 56	75.5	72.3	73.2	74.7	0.478		80.0	62.4	68.0	75.7	0.424
	≥ 64	58.0	83.6	78.0	66.6	0.416		59.4	77.9	72.9	65.7	0.373
	≥72	31.3	94.9	85.9	58.0	0.262		43.9	88.2	78.8	61.1	0.321
	≥ 80	14.4	97.9	87.1	53.3	0.122		23.3	97.0	88.5	55.8	0.203
	≥ 88	2.7	99.9	96.2	50.7	0.026		7.0	99.7	95.8	99.7	0.067
	≥96	0.3	100	100	50.1	0.003		1.2	100	100	50.3	0.012
	AUC (95%CI): 0.81(0.80 – 0.82); p<0.001							AUC (95%CI): 0.78(0.74 – 0.81); p<0.001
Standardized weighting [10*βw]	Cut off for the risk score	Se%	Sp%	PPV %	NPV %	K		Se%	Sp%	PPV %	NPV %	K
	≥ 8	99.7	1.5	50.3	82.4	0.012		100.0	0.9	50.2	100	0.009
	≥ 16	98.9	10.5	52.5	90.8	0.095		98.5	7.6	51.6	83.3	0.061
	≥ 24	98.5	17.4	54.4	92.1	0.159		97.6	12.4	52.7	83.7	0.100
	≥ 32	97.8	23.8	56.2	91.4	0.215		97.0	19.7	54.7	86.7	0.167
	≥ 40	93.4	41.4	61.5	86.3	0.349		93.9	33.9	58.7	84.8	0.279
	≥ 48	87.5	58.6	67.9	82.4	0.461		90.0	48.5	63.6	82.9	0.385
	≥ 56	77.6	71.9	73.4	76.3	0.495		81.5	63.9	69.3	77.6	0.455
	≥ 64	60.7	82.5	77.6	67.7	0.432		58.2	78.8	73.3	65.3	0.370
	≥72	38.3	93.2	84.9	60.2	0.314		38.8	89.7	79.0	59.4	0.285
	≥ 80	16.1	97.7	87.2	53.8	0.137		22.7	97.3	89.3	55.7	0.200
	≥ 88	4.3	99.7	93.0	51.0	0.040		4.8	99.7	94.1	51.2	0.045
	≥96	0.3	50.1	100	50.0	0.003		0.6	100	100	50.2	0.006
	AUC (95%CI): 0.81(0.80–0.83); p<0.001							AUC (95%CI): 0.78(0.74 – 0.81); p<0.001

S_e-Sensitivity; S_p-Specificity; PPV-Positive Predictive Value; NPV-Negative Predictive Value; K-Cohen’s Kappa; AUC-Area Under the ROC; CI-Confidence Interval.

Fig. 2.

ROC Curves for training and testing dataset for Ischemic strokes using constant and standardised weighting.

3.3. Risk score model for occurrence of hemorrhagic strokes

Among 937 hemorrhagic stroke cases and 937 matching controls, 7 risk factors were included in the final model using the training set (Supplementary Table S4). These include income <$100, hypertension, dyslipidemia, diabetes mellitus, raised waist-hip ratio, sprinkling of salt on food at table and low consumption of green leafy vegetables. The statistical weights and classification of stroke cases and controls using the scoring systems are shown in Tables S5 and S6. The validity characteristics of the risk scoring system at various cut-off are shown in Table 6. The Cohen’s kappa was maximal at a total risk score of 63% for both the constant and the standardized weighting methods with AUCs of 0.78, p<0.001 or 0.78, p<0.001 (Figure 3) for the test dataset, depending on the weighting approach used.

Table 6:

Performance of the risk score for Hemorrhagic stroke types using training and test dataset

		Training Dataset					Test Dataset
	Cut off for the risk score	Se%	Sp%	PPV %	NPV %	K	Se%	Sp%	PPV %	NPV %	K
Constant weighting [10*β]	≥ 9	99.5	11.4	52.9	11.4	0.109	97.9	9.9	52.1	82.4	0.078
	≥ 18	98.9	28.4	58.0	96.2	0.273	97.2	26.2	56.8	90.2	0.234
	≥ 27	98.4	41.1	62.5	96.2	0.395	96.5	39.0	61.3	91.7	0.355
	≥ 36	98.2	45.6	64.4	96.3	0.439	96.5	48.9	65.4	93.2	0.454
	≥ 45	98.0	46.7	64.8	95.9	0.447	96.5	49.6	65.7	93.3	0.461
	≥ 54	95.6	50.3	65.8	92.0	0.459	92.2	52.5	66.0	87.1	0.447
	≥ 63	83.5	66.4	71.3	80.1	0.499	87.9	61.7	69.7	83.7	0.496
	≥ 72	49.9	84.7	76.5	62.8	0.346	69.5	70.2	70.0	69.7	0.397
	≥81	16.8	97.2	85.9	53.9	0.140	47.5	85.1	76.1	61.9	0.326
	≥ 90	1.4	100	100	50.3	0.014	8.5	98.6	85.7	51.9	0.071
	≥ 99	0.1	100	100	50.0	0.001	2.0	100	100	50.5	0.021
	AUC (95%CI): 0.81(0.79 – 0.83); p<0.001						AUC (95%CI): 0.78(0.73 – 0.83);p<0.001
Standardized weighting [10*βw]	Cut off for the risk score	Se%	Sp%	PPV %	NPV %	K	Se%	Sp%	PPV %	NPV %	K
	≥ 9	99.5	7.6	51.9	93.8	0.071	97.9	10.6	52.3	83.3	0.085
	≥ 18	98.9	24.3	56.6	95.6	0.232	97.2	21.3	55.2	88.2	0.184
	≥ 27	98.5	36.2	60.7	96.0	0.347	96.5	39.0	61.3	91.7	0.355
	≥ 36	98.2	45.6	64.4	96.3	0.439	96.5	46.1	64.2	92.9	0.426
	≥ 45	98.0	46.7	64.8	95.9	0.447	96.5	49.6	65.7	93.3	0.461
	≥ 54	96.4	49.4	65.6	93.1	0.457	92.2	52.5	66.0	87.1	0.447
	≥ 63	87.0	62.5	69.9	82.8	0.495	87.9	61.7	69.7	83.7	0.496
	≥ 72	57.0	81.2	75.2	65.4	0.382	69.5	70.9	70.5	69.9	0.404
	≥81	28.3	93.4	81.0	56.6	0.217	39.7	87.2	75.7	59.1	0.270
	≥ 90	3.4	99.5	87.1	50.7	0.029	8.5	98.6	85.7	51.9	0.071
	≥ 99	0.1	100	100	50.0	0.001	2.1	100	100	50.5	0.021
	AUC (95%CI): 0.81(0.79–0.83); p<0.001						AUC (95%CI): 0.79(0.73 – 0.84); p<0.001

S_e-Sensitivity; S_p-Specificity; PPV-Positive Predictive Value; NPV-Negative Predictive Value; K-Cohen’s Kappa; AUC-Area Under the ROC; CI-Confidence Interval.

Fig. 3.

ROC Curves for training and testing dataset for Hemorrhagic strokes using constant and standardised weighting.

3.4. Risk score models for occurrence of strokes by sex

The aggregate risk score information by sex for 1,938 male (Tables S7 to S10) and 1,615 female (Tables S11 to S14) case-control pairs are shown in the supplementary data section.

4.0. DISCUSSION

This is the first effort aimed at developing a stroke risk scoring instrument for indigenous Africans. Using a case-control methodology among 3,553 pairs of stroke cases and stroke-free controls, our method was able to identify 13 potentially modifiable (income level <$100, education, hypertension, dyslipidemia, diabetes mellitus, cardiac disease, raised waist-hip ratio, stress, sprinkling salt on food at table, low-consumption of green leafy vegetable, regular sugar consumption, physical inactivity and regular meat consumption) and two non-modifiable (advancing age and family history of CVD) risk factors associated with occurrence of stroke. Due to imprecise matching of age, we included age as a covariate in our models. By assigning weights to these individual risk factors according to their β coefficients obtained from multivariate logistic regression models, hypertension, dyslipidemia, diabetes and low consumption of green leafy vegetables were found to be the top four modifiable risk factors which attracted the highest weights with stress having the lowest weight using both constant and standardized weighting approaches. The rationale for ascribing weights to risk factors is similar to how regression coefficients obtained from Cox regression hazards for prospective studies are utilized in the development of risk prediction models.^2–10

In the present study, when the test dataset was used to assess the performance of our scoring system, the overall predictive accuracy of the risk scoring instrument (for the occurrence of stroke taking into account 15 risk factors), had a predictive probability measured using Area Under Curve (AUC) of 76% (95% CI of 73 to 79%) or 76% (95% CI of 74 to 79%) depending on the weighting approach used. Similar diagnostic properties were obtained for ischemic stroke models using 12 risk factors and 7 risk factors for hemorrhagic stroke with AUC of 78% (95%CI: 74 to 81%) for Ischemic stroke (irrespective of weighting approach) and 785% (95%CI: 73% to 83%) or 79% (95%CI: 73 to 84%) for Hemorrhagic stroke, depending on the weighting approach used. By far, the most popular stroke risk prediction model is the Framingham Stroke Risk Score (FSRS) which incorporates systolic blood pressure, diabetes mellitus, cigarette smoking, prior cardiovascular disease, atrial fibrillation, left ventricular hypertrophy and use of antihypertensive medications.^1,23 The AUC of the factors in the FSRS for prediction of stroke have been fair to good, ranging from 65% (95% CI: 62 – 67%) from the Atherosclerosis Risk in Communities (ARIC) cohort,²⁴ 60% (95% CI: 58 – 63%) for the Cardiovascular Health Study (CHS)²⁵, 71% (95%CI: 67 – 75%) from the Framingham Heart Study (FHS)²⁶ and 59% (95%CI:57 – 61%) the first cohort of the Rotterdam Study (RS)²⁷. A meta-analysis shows that the AUC for the 7 factors in the FSRS from these four cohorts was 62% (95% CI: 61 – 63%).²⁸ However, it should be emphasized that these AUCs for the FSRS are not directly comparable with those of ours due to differences in study design and methodology used to derive its predictive accuracy. For this comparison, an Indian study involving hemorrhagic stroke cases and controls (n=166) using a similar analytic approach using 5 factors namely hypertension, raised serum total cholesterol, use of anticoagulants and antiplatelet agents, past history of transient ischemic attack and alcohol had a predictive accuracy of 79% with a similar 95% confidence interval of 73 to 84%.²⁹

There was however an appreciable overlap in cut-off scores in discriminating between cases and controls (Tables 3, 5 and 6) in our study. Thus, in a broader sense the discriminatory properties based on area-under curve plots were ‘good’ but not excellent and may require further refinements. For instance, an item such as hypertension was analyzed at ‘yes’ or ‘no’ but could be further categorized based on severity of blood pressure values to increase the discriminatory quality of this ubiquitous risk factor present among stroke and stroke-free subjects. Furthermore, dyslipidemia, which is a composite term defined in this study as elevations in total cholesterol, LDL-cholesterol, triglyceride concentration or decrement in HDL-cholesterol may have to be dis-aggregated and tested individually or as ratios or using different cut-off values for lipid sub-fractions to further separate cases from controls. Again, dietary factors such as sprinkling of salt on food at table, low consumption of green vegetable, regular sugar consumption and meat consumption may be constructed into a risk dietary score which considers tiers of regularity of dietary practices. An advantage with the approach we have adapted in developing this risk scoring system is in its good predictive accuracy of about 79% and its basis on a broad range of putative risk factors associated with stroke occurrence in Africa. Furthermore, as posited by Herman and colleagues³⁰, we deployed a somewhat back-validation method by splitting the same dataset to training and test sets to validate our risk scoring system. Admittedly our approach is unconventional in the development of risk score, however there are instances in literature where such an approach has been utilized.^15,31–35 Going forward, it would be ideal to validate this in other dataset available on the continent such as the Cardiovascular H3Africa Innovation Resource (CHAIR) resource from the H3Africa networks.^36,37

Despite the advantages, and possible improvements suggested, further validation is required using community dwelling, stroke-free, prospective cohorts. At the moment, there are no such longitudinal studies on-going in Africa largely due to limited funding available to undertake such an ambitious research projects as done in several high-income countries. The literature shows that some of these studies have accrued data from electronic health records and insurance claims records which are available in some African countries. Ascertainment bias for some risk factors such as stress, dietary history, smoking and alcohol intake might have been present in the stroke cases because those who were aphasic or unconscious on admission had these risk factors assessed using reliable proxies. We also acknowledge issues of selection bias as a concern in the use of case-control studies for predictive models. These challenges notwithstanding, this is the first attempt at developing an Afrocentric stroke risk scoring scale.

The information provided from these models are intended to be utilized for primary prevention drive for stroke and its primary types on the continent. Ultimately, the SIREN Afrocentric risk score can be further enhanced by integration of polygenic risk score that incorporates aggregate genetic risk from single nucleotide polymorphisms associated with stroke. A study which looked at incorporating 324 SNPs associated with stroke and its risk factors resulted in small improvement in prediction of future stroke using the Framingham Stroke Risk Score compared to the classical epidemiological risk factors for stroke.²⁸. The enhancement in our own population may be higher due to the higher heritability of stroke in people of African ancestry.³⁸

5.0. CONCLUSION

We have developed aggregate risk assessment scores for stroke and its primary types in indigenous Africans. Population-based prospective cohort studies are required to refine and validate this novel tool. This will facilitate risk factor assessment for primary prevention of stroke and indeed other cardiovascular diseases in Africa at the population level.

HIGHLIGHTS.

• No Afro-centric aggregate stroke risk scoring instrument is available for indigenous Africans

• Using data from the SIREN study, we developed a novel stroke riskometer for indigenous Africans

• The instrument included 15 risk factors for stroke occurrence, 12 for Ischemic stroke and 7 for hemorrhagic stroke.

• It had a predictive accuracy of 76%, sensitivity of 80% and specificity of 63%

• Prospective studies are required to further validate this novel instrument for population-wide deployment to boost stroke prevention in Africa