Exploring the potential and limitations of deep learning and explainable AI for longitudinal life course analysis

Coupland, Helen; Scheidwasser, Neil; Katsiferis, Alexandros; Davies, Megan; Flaxman, Seth; Hulvej Rod, Naja; Mishra, Swapnil; Bhatt, Samir; Unwin, H. Juliette T.

doi:10.1186/s12889-025-22705-4

BMC Public Health

Table 2 Relative model performance within four LCPs; period, repeats, order, and timing

From: Exploring the potential and limitations of deep learning and explainable AI for longitudinal life course analysis

	LR				XGBoost				MLSTM-FCN				InceptionTime				ResNet				LSTMAttention
	F1	PRC	AUC	Brier	F1	PRC	AUC	Brier	F1	PRC	AUC	Brier	F1	PRC	AUC	Brier	F1	PRC	AUC	Brier	F1	PRC	AUC	Brier
Period
1	100.00	100.00	100.00	0.07	99.92	100.00	100.00	0.03	100.00	100.00	100.00	4.10	99.92	99.99	100.00	0.25	99.89	99.94	99.95	0.12	100.00	100.00	100.00	1.05
2	99.82	99.96	100.00	0.25	99.93	100.00	100.00	0.02	99.67	99.93	100.00	0.03	99.71	99.84	99.92	0.06	99.82	99.95	100.00	0.09	100.00	100.00	100.00	0.01
3	81.09	89.26	99.69	0.50	97.15	99.22	99.97	0.12	91.91	95.47	99.88	1.17	95.96	97.65	99.97	0.21	98.61	99.73	99.99	0.06	98.47	99.89	100.00	0.06
Repeat
1	50.42	51.52	98.23	0.92	99.45	99.98	100.00	0.04	99.82	99.99	100.00	0.06	99.33	99.84	99.99	0.06	99.51	99.95	100.00	0.05	99.94	100.00	100.00	0.01
2	39.38	30.78	95.91	1.08	98.28	99.68	100.00	0.05	97.84	98.67	99.95	0.14	96.96	97.06	99.92	1.37	97.88	98.45	99.39	0.10	96.37	97.15	99.97	0.14
3	42.29	36.87	97.53	1.40	90.07	95.81	99.92	0.33	99.67	99.87	99.99	0.05	96.88	98.68	99.97	3.93	98.28	99.38	99.99	0.10	100.00	100.00	100.00	0.02
Order
1	57.48	44.93	95.67	4.20	97.70	98.20	99.93	0.37	99.04	99.39	99.63	2.64	99.80	99.99	100.00	2.48	99.76	99.95	100.00	0.07	98.89	99.80	99.99	0.20
2	34.57	24.72	95.94	1.77	86.39	90.63	99.79	0.47	95.48	97.50	99.44	0.20	98.82	99.84	99.99	0.05	97.46	98.97	99.82	0.16	99.37	99.96	100.00	0.04
3	27.24	17.12	93.75	1.70	91.81	94.34	99.88	0.35	98.60	98.89	99.37	0.08	99.58	99.98	100.00	0.02	99.35	99.79	99.92	0.03	99.86	99.90	99.99	0.01
Timing
1	36.49	26.59	93.47	2.67	99.73	99.99	100.00	0.06	99.22	99.95	100.00	0.83	99.89	99.99	100.00	0.07	100.00	100.00	100.00	0.43	100.00	100.00	100.00	0.01
2	28.26	21.15	95.76	0.99	95.03	98.71	99.98	0.16	97.01	99.40	99.99	0.06	95.77	98.68	99.95	0.42	97.98	99.70	99.99	0.08	77.01	75.01	99.47	1.81
3	54.40	44.87	98.17	1.83	93.81	98.15	99.92	0.31	97.86	99.21	99.84	0.16	99.24	99.80	99.97	0.04	98.97	99.67	99.99	0.06	65.41	59.82	97.54	25.05

Each row gives information about a rule including the LCP it falls into, an identifying number, the proportion of the resulting outcome that is positive and its definition. If the rule is met, then the outcome is positive. For each model the F1, AUROC and AUPRC values as percentages are provided and the highest performing model for each dataset is highlighted in bold font. 10 percent noise in the data. LR = logistic regression. Note F1 scores depend on the chosen threshold which in this case was selected to maximise the F1 score. All other metrics are independent of the threshold

Back to article page

ISSN: 1471-2458

Contact us

General enquiries: journalsubmissions@springernature.com